Cyanobacterial bloom evolution prediction method and system based on improved predrnn for a watershed

Abstract

The present application relates to the technical field of recurrent neural network prediction, in particular to a watershed blue-green algae bloom evolution prediction method and system based on improved PredRNN. The specific implementation process comprises: constructing the spatio-temporal feature sequence of the watershed; in the PredRNN, the spatial position migration is extracted by using the flow field flow velocity vector, the distribution intensity fluctuation is extracted by using the environmental factor, and the mixed hidden state is generated by feature recoupling; based on the cross-step history recall mechanism, the historical evolution trend item is retrieved from the watershed historical evolution library, the mixed hidden state is information compensated through the memory enhancement channel, and the water bloom evolution distribution field of the future period is output. Through the fusion of hydrodynamic constraint, environmental modulation mechanism and historical memory enhancement, the present application effectively improves the problem of insufficient description of migration mechanism in water bloom prediction, long-time sequence prediction easy to blur and distorted, and improves the spatial retention and time stability of watershed blue-green algae bloom evolution prediction.

Description

Technical Field

[0001] This invention relates to the field of predictive recurrent neural network technology, specifically to a method and system for predicting the evolution of cyanobacterial blooms in watersheds based on an improved PredRNN. Background Technology

[0002] The evolution of cyanobacterial blooms in watersheds is a typical non-stationary spatiotemporal dynamic process, characterized by its sudden and rapid state distribution, as well as its complex spatial topology, posing a severe challenge to aquatic ecological security. Therefore, it is of great significance to use specific computational models to predict the future state evolution trend of this complex system. With the development of artificial intelligence and deep learning technologies, existing techniques typically employ neural network architectures (such as ConvLSTM networks) that combine convolutional operations and sequence memory mechanisms to handle such problems. These methods aim to capture spatial structural information and temporal evolution dependencies in historical monitoring data sequences by constructing recurrent memory units within the neural network, thereby generating the future state distribution of the bloom through model inference.

[0003] However, the existing technologies based on the basic recursive architecture have inherent limitations in long-term time-series prediction tasks. For standard ConvLSTM computational units, although local perception is introduced through convolution operations, the information retention rate of the memory stream often decreases with the increase of the prediction step size when transferring hidden layer states along the time axis, resulting in insufficient long-range dependency capture ability (i.e., gradient vanishing). The state transition mechanism of existing models is difficult to distinguish between translational motion features dominated by hydrodynamics and distribution intensity features dominated by evolutionary fluctuations. This feature coupling leads to severe information dissipation in the high-dimensional feature representation of the output model after multi-step iterative inference. The generated prediction data field loses high-frequency detail information, exhibiting a phenomenon where the prediction results tend to be averaged (fuzzy), which cannot meet the needs of refined state prediction. In summary, existing technologies cannot simultaneously decouple and characterize the efficient spatiotemporal memory mechanism of two heterogeneous dynamic features of cyanobacterial bloom physical drift and distribution intensity. As a result, existing models suffer from spatial feature fuzziness due to gradient loss during long-term extrapolation, making it impossible to achieve high-fidelity prediction of the evolution of highly dynamic, non-rigid cyanobacterial blooms.

[0004] To address this, a method and system for predicting the evolution of cyanobacterial blooms in watersheds based on an improved PredRNN were proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the evolution of cyanobacterial blooms in watersheds based on an improved PredRNN, so as to predict the evolution of cyanobacterial blooms in watersheds.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A watershed cyanobacterial bloom evolution prediction method based on improved PredRNN includes:

[0008] The monitoring data of chlorophyll concentration, flow field velocity vector and environmental factors in the target watershed are acquired and spatiotemporally aligned and interpolated in a gridded manner. The spatiotemporal feature sequence of algal bloom is constructed and input into PredRNN.

[0009] In PredRNN, a Lagrange translational flow convolution gate, which includes displacement guidance, dynamic sampling, and gated fusion, is used to extract spatial position migration constrained by the flow field velocity vector; a reaction-diffusion wave gate, which includes environmental modulation, wave update, and diffusion constraint, is used to extract distribution intensity fluctuations regulated by environmental factors.

[0010] Spatial location migration and distribution intensity fluctuations are independently evolved and updated in an orthogonal subspace, and converged into a mixed latent state of the watershed at the current moment through feature recoupling; using a cross-step historical recall mechanism, the topological similarity between the mixed latent state and each memory anchor point in the watershed historical evolution database is calculated, and historical evolution trend items that conform to chlorophyll concentration are retrieved and aggregated.

[0011] By inputting historical evolution trend terms into the memory enhancement pathway, nonlocal information supplementation and uncertainty compensation are provided for the mixed latent states, and the distribution field of algal bloom evolution in future periods is predicted.

[0012] Preferably, the specific implementation process of acquiring monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performing spatiotemporal alignment and gridded interpolation, and constructing a spatiotemporal feature sequence of algal blooms for input into PredRNN includes:

[0013] A multi-source heterogeneous sensor array was used to collect chlorophyll fluorescence intensity values ​​at monitoring sections, flow velocity vector components measured by acoustic Doppler current profilers, and environmental parameters such as light and temperature. Outliers were removed using statistical filtering algorithms, and missing data were filled using neighborhood mean to form the original observation dataset. Based on a unified time reference, the original observation dataset was resampled at frequency and synchronized with temporal phase to eliminate time delays between multi-source heterogeneous data. A regularized Euclidean grid covering the entire watershed was constructed, and a spatial interpolation algorithm was used to map the original observation datasets of discrete monitoring stations to the corresponding grid nodes, generating a continuous spatial raster containing chlorophyll concentration, flow velocity vector, and environmental factors. The continuous spatial raster was subjected to max-min dimensionless normalization, and chlorophyll concentration, flow velocity vector, and environmental factors were stacked along the channel dimension to construct a multidimensional state feature tensor. The multidimensional state feature tensor was extracted in continuous time sequence using a sliding window according to the temporal evolution order and combined to form a spatiotemporal feature sequence of algal bloom.

[0014] Preferably, in PredRNN, the specific implementation process for extracting spatial position migration constrained by the flow field velocity vector through a Lagrange translation flow convolution gate that includes displacement guidance, dynamic sampling, and gated fusion includes:

[0015] The Lagrange translational flow convolution gate jointly maps the input spatiotemporal feature sequence of algal bloom with the flow field velocity vector of the corresponding time slice to generate displacement guidance information consistent with the spatial grid coordinates. Based on the displacement guidance information, the sampling coordinates of the feature tensors representing the spatial distribution in the spatiotemporal feature sequence of algal bloom are continuously offset and mapped to form dynamic sampling features that change with the flow field. The dynamic sampling features are input into the convolution operation path, and the dynamic sampling features are weighted and integrated with the current hidden state of PredRNN through gating fusion to output the spatial position migration of cyanobacteria that changes with the flow field.

[0016] Preferably, the specific implementation process for extracting distribution intensity fluctuations modulated by environmental factors through a reaction-diffusion fluctuation gate that includes environmental modulation, fluctuation update, and diffusion constraint includes:

[0017] The reaction-diffusion fluctuation gate jointly maps the feature channels representing chlorophyll concentration changes in the spatiotemporal feature sequence of algal blooms with environmental factors of the corresponding time slices to generate environmental modulation information corresponding to the feature channels of the hidden state. Based on the environmental modulation information, the feature components related to algal bloom intensity changes in the hidden state of PredRNN are updated with modulated temporal difference to form fluctuation intermediate features. The fluctuation intermediate features are subjected to diffusion constraint processing with temporal continuity, and the output is the distribution intensity fluctuation.

[0018] Preferably, the specific implementation process of independently evolving and updating spatial location migration and distribution intensity fluctuations within an orthogonal subspace, and then converging them into a hybrid latent state of the watershed at the current moment through feature recoupling includes:

[0019] The spatial position migration output by the Lagrange translation flow convolution gate and the distribution intensity fluctuation output by the reaction diffusion wave gate are respectively mapped to mutually orthogonal feature subspaces. PredRNN state updates are performed along the time dimension in their respective subspaces, and independent temporal evolution is maintained. At the current time, the state representations in the two subspaces are aligned and fused by the feature recoupling operator to generate a hybrid hidden state that simultaneously contains spatial migration information and distribution intensity information.

[0020] Preferably, the specific implementation process of using a cross-step-length history recall mechanism to calculate the topological similarity between the mixed hidden state and each memory anchor point in the watershed history evolution database, and to retrieve and aggregate historical evolution trend items that conform to chlorophyll concentration includes:

[0021] The cross-step historical recall mechanism inputs the current mixed latent state into the historical recall path and aligns it with the multi-timescale memory anchors stored in the watershed historical evolution database. Based on the joint representation of the mixed latent state in the dimensions of spatial structure distribution and algal bloom intensity change, it calculates the topological similarity between the mixed latent state and each memory anchor. The topological similarity results are used to filter historical evolution segments that match the current chlorophyll concentration, and the historical evolution segments are weighted and aggregated to output the historical evolution trend term.

[0022] Preferably, the specific implementation process of inputting historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed latent states, and predicting the distribution field of algal bloom evolution in future periods includes:

[0023] The historical evolution trend term is jointly mapped with the current mixed hidden state to form an enhanced state representation, which is then input into the memory enhancement pathway. In the memory enhancement pathway, nonlocal information transfer across spatial regions is performed to supplement features in regions with sparse and insufficient information in the mixed hidden state, and the uncertainty in the state representation is compensated and adjusted based on the historical evolution trend term. The enhanced state representation is then input into the PredRNN output pathway to generate the algal bloom evolution distribution field corresponding to the future time period.

[0024] A watershed cyanobacterial bloom evolution prediction system based on an improved PredRNN includes:

[0025] The data processing module acquires monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performs spatiotemporal alignment and gridded interpolation, and constructs a spatiotemporal feature sequence of algal blooms, which is then input into PredRNN.

[0026] In the feature extraction module of PredRNN, spatial position migration constrained by flow field velocity vector is extracted through a Lagrange translational flow convolution gate that includes displacement guidance, dynamic sampling and gated fusion; and distribution intensity fluctuations that are regulated by environmental factors are extracted through a reaction-diffusion fluctuation gate that includes environmental modulation, fluctuation update and diffusion constraint.

[0027] The historical recall module independently evolves and updates the spatial location migration and distribution intensity fluctuations within the orthogonal subspace, and converges them into a mixed hidden state of the watershed at the current moment through feature recoupling; using the cross-step historical recall mechanism, it calculates the topological similarity between the mixed hidden state and each memory anchor point in the watershed historical evolution database, and retrieves and aggregates historical evolution trend items that conform to chlorophyll concentration.

[0028] The evolution prediction module inputs historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed latent state, and predicts the distribution field of algal bloom evolution in future periods.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. This invention introduces a Lagrange translational flow convolution gate into PredRNN to explicitly embed the flow field velocity vector into the spatiotemporal feature extraction process, thereby realizing the physical constraint modeling of the spatial location migration of cyanobacterial blooms caused by water flow. This effectively improves the problem that ConvLSTM or traditional PredRNN methods are difficult to characterize translation-dominated spatial evolution, and enhances the prediction accuracy of the drift path and morphological changes of watershed-scale blooms.

[0031] 2. This invention constructs a reaction-diffusion wave gate that includes environmental modulation, wave update, and diffusion constraints, thereby decoupling the distribution intensity of cyanobacterial bloom evolution from changes in environmental factors. This enables the model to characterize the evolutionary characteristics of bloom intensity dominated by evolutionary waves, avoiding the information dissipation problem caused by mixed modeling of physical drift and distribution intensity, thus improving the ability to characterize sudden changes and non-stationary evolution processes of blooms.

[0032] 3. This invention effectively alleviates the gradient decay and memory loss problems in long-term time series prediction by synergistic design of independent evolution of orthogonal subspaces, cross-step-length history recall mechanism and memory enhancement pathway, improves the preservation of spatial details of algal bloom evolution distribution field and the stability of long-term extrapolation, and can meet the needs of watershed cyanobacterial blooms for refined and high-fidelity forecasting. Attached Figure Description

[0033] Figure 1 This is a flowchart of the watershed cyanobacterial bloom evolution prediction method based on improved PredRNN proposed in this invention;

[0034] Figure 2 This is a structural diagram of the watershed cyanobacterial bloom evolution prediction system based on improved PredRNN proposed in this invention;

[0035] Figure 3 This is a schematic diagram of the watershed cyanobacterial bloom evolution prediction process proposed in this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It must be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to constitute any limitation on the scope of protection of this invention. Therefore, all equivalent changes or modifications conceived by those skilled in the art based on the content disclosed in this invention without inventive effort should fall within the scope of protection claimed by this invention.

[0037] Reference Figures 1 to 3This invention proposes a method and system for predicting the evolution of cyanobacterial blooms in watersheds based on an improved PredRNN. The technical solution is as follows:

[0038] Example 1:

[0039] Reference Figure 1 This embodiment proposes a watershed cyanobacterial bloom evolution prediction method based on an improved PredRNN, including:

[0040] The monitoring data of chlorophyll concentration, flow field velocity vector and environmental factors in the target watershed are acquired and spatiotemporally aligned and interpolated in a gridded manner. The spatiotemporal feature sequence of algal bloom is constructed and input into PredRNN.

[0041] In PredRNN, a Lagrange translational flow convolution gate, which includes displacement guidance, dynamic sampling, and gated fusion, is used to extract spatial position migration constrained by the flow field velocity vector; a reaction-diffusion wave gate, which includes environmental modulation, wave update, and diffusion constraint, is used to extract distribution intensity fluctuations regulated by environmental factors.

[0042] Spatial location migration and distribution intensity fluctuations are independently evolved and updated in an orthogonal subspace, and converged into a mixed latent state of the watershed at the current moment through feature recoupling; using a cross-step historical recall mechanism, the topological similarity between the mixed latent state and each memory anchor point in the watershed historical evolution database is calculated, and historical evolution trend items that conform to chlorophyll concentration are retrieved and aggregated.

[0043] By inputting historical evolution trend terms into the memory enhancement pathway, nonlocal information supplementation and uncertainty compensation are provided for the mixed latent states, and the distribution field of algal bloom evolution in future periods is predicted.

[0044] Furthermore, the specific implementation process of acquiring monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performing spatiotemporal alignment and gridded interpolation, and constructing the spatiotemporal feature sequence of algal blooms as input to PredRNN includes:

[0045] A multi-source heterogeneous sensor array was used to collect chlorophyll fluorescence intensity values ​​at monitoring sections, flow velocity vector components measured by acoustic Doppler current profilers, and environmental parameters such as light and temperature. Outliers were removed using statistical filtering algorithms, and missing data were filled using neighborhood mean to form the original observation dataset. Based on a unified time reference, the original observation dataset was resampled at frequency and synchronized with temporal phase to eliminate time delays between multi-source heterogeneous data. A regularized Euclidean grid covering the entire watershed was constructed, and a spatial interpolation algorithm was used to map the original observation datasets of discrete monitoring stations to the corresponding grid nodes, generating a continuous spatial raster containing chlorophyll concentration, flow velocity vector, and environmental factors. The continuous spatial raster was subjected to max-min dimensionless normalization, and chlorophyll concentration, flow velocity vector, and environmental factors were stacked along the channel dimension to construct a multidimensional state feature tensor. The multidimensional state feature tensor was extracted in continuous time sequence using a sliding window according to the temporal evolution order and combined to form a spatiotemporal feature sequence of algal bloom.

[0046] Specifically, multiple types of monitoring sections are deployed along the main river channel and key tributaries within the target watershed, and long-term continuous observations are conducted synchronously using a multi-source heterogeneous sensor array. In this embodiment, chlorophyll concentration data is acquired by an online chlorophyll fluorescence sensor with a sampling period set to the minute level to characterize the rapid changes in the intensity of cyanobacterial blooms; the flow field velocity vector is acquired by an acoustic Doppler profile current meter to obtain two-dimensional velocity information including the mainstream component and the transverse component, with a time resolution lower than that of the chlorophyll monitoring frequency; environmental factor data are acquired by meteorological and water quality monitoring equipment, including parameters that affect the evolution of cyanobacterial blooms such as water temperature and light intensity.

[0047] The multi-source heterogeneous sensors are inevitably affected by factors such as communication jitter, equipment aging, and local interference during long-term operation. Therefore, before generating data that can be used as model input, the raw monitoring data undergoes quality control processing. This quality control includes performing statistical filtering on the monitoring sequence to identify and remove outliers that significantly deviate from the historical distribution, and performing mean compensation processing on short-term missing segments based on valid observations within the temporal or spatial neighborhood, thereby forming a continuous and stable raw observation dataset. This processing ensures that the input data maintains statistical continuity and consistency, preventing abnormal noise from being mislearned during subsequent neural network training.

[0048] Because the sampling frequencies and timestamps for chlorophyll concentration, flow velocity vector, and environmental factors are inconsistent, it is necessary to perform frequency resampling and temporal phase synchronization processing on the original observation dataset based on a unified time reference. Specifically, the time step of the prediction model is used as a unified time axis. High-frequency data is downsampled, and low-frequency data is interpolated to ensure that corresponding chlorophyll concentration, flow velocity vector, and environmental factor information can be obtained at each time slice. This eliminates the time delay error caused by multi-source heterogeneous sampling and improves the alignment accuracy of temporal features. After completing the temporal dimension alignment, to address the problem of discrete spatial distribution in the watershed, this embodiment constructs a regularized Euclidean grid covering the entire target watershed. The grid divides the water area with a fixed spatial resolution, and the discrete observation values ​​obtained from each monitoring section are mapped to the corresponding grid nodes through spatial interpolation. This generates a spatially continuous chlorophyll concentration field, flow velocity vector field, and environmental factor field, transforming the originally irregularly distributed monitoring data into a rasterized representation suitable for convolutional and recurrent neural network processing.

[0049] To eliminate the impact of differences in the dimensions of different physical quantities on the model training process, the continuous spatial grid undergoes dimensionless processing before being input into the PredRNN. Specifically, the data of each channel is subjected to max-min normalization, so that chlorophyll concentration, flow velocity vector components, and environmental factors are all mapped to a unified numerical range, thereby preventing any one type of feature from dominating the model parameter update due to an excessively large numerical range. Subsequently, the normalized features are stacked along the channel dimension to form a multidimensional state feature tensor that can simultaneously characterize the intensity of cyanobacterial distribution, hydrodynamic characteristics, and environmental driving factors. In the time dimension, according to the natural temporal sequence of algal bloom evolution, a fixed-length time segment is extracted from the multidimensional state feature tensor of continuous time using a sliding window method, and multiple adjacent time segments are combined into an algal bloom spatiotemporal feature sequence. This feature sequence serves as the input to the PredRNN, enabling the model to learn both spatial structure changes and temporal evolution patterns during the recursive state update process.

[0050] This embodiment achieves a unified representation of multi-source heterogeneous hydrological, water quality, and environmental data at the data level, effectively improving the problem of unstable model convergence caused by data time asynchrony, high spatial dispersion, and large dimensional differences. The spatiotemporal feature sequence of algal blooms constructed in this way provides a high-quality input foundation for the decoupled modeling of spatial migration and distribution intensity in the subsequent PredRNN, enabling the model to maintain higher spatial consistency and prediction accuracy in long-term time series prediction tasks, thereby improving the reliability of watershed cyanobacterial bloom evolution forecasting.

[0051] Furthermore, in PredRNN, the specific implementation process for extracting spatial position migration constrained by the flow field velocity vector through a Lagrange translation flow convolution gate that includes displacement guidance, dynamic sampling, and gated fusion includes:

[0052] The Lagrange translational flow convolution gate jointly maps the input spatiotemporal feature sequence of algal bloom with the flow field velocity vector of the corresponding time slice to generate displacement guidance information consistent with the spatial grid coordinates. Based on the displacement guidance information, the sampling coordinates of the feature tensors representing the spatial distribution in the spatiotemporal feature sequence of algal bloom are continuously offset and mapped to form dynamic sampling features that change with the flow field. The dynamic sampling features are input into the convolution operation path, and the dynamic sampling features are weighted and integrated with the current hidden state of PredRNN through gating fusion to output the spatial position migration of cyanobacteria that changes with the flow field.

[0053] Specifically, the Lagrange translational flow convolution gate is deployed in the state update path of PredRNN, and its input includes the spatiotemporal feature sequence of the algal bloom at the current moment and the flow field velocity vector data corresponding to the time slice. The spatiotemporal feature sequence of the algal bloom is a multi-channel feature tensor after spatiotemporal alignment and gridding; the flow field velocity vector is a two-dimensional velocity component that corresponds one-to-one with the spatial grid, used to characterize the dominant flow direction and intensity of the water body within the time slice.

[0054] The Lagrange translational flow convolution gate first performs a joint mapping operation, fusing and encoding the spatiotemporal feature sequence of the algal bloom with the flow field velocity vector in the channel dimension to generate displacement guidance information strictly aligned with the spatial grid coordinates. Specifically, the joint mapping first concatenates and stacks the spatiotemporal feature sequence of the algal bloom with the flow field velocity vector in the channel dimension, and then fuses and projects the physical velocity with the nonlinear motion trend implied by the features, outputting a dense tensor with two channels as displacement guidance information. This displacement guidance information has the same spatial resolution as the input grid in terms of data structure, and its two channels represent the floating-point coordinate offsets that the material at each grid point needs to move in the horizontal and vertical directions, respectively. Based on this, the specific process of the continuous offset mapping is as follows: first, a standard regular grid with the same size as the current feature tensor is constructed, where each node stores its own integer coordinate index; then, the coordinate offsets in the displacement guidance information are added element by element to the corresponding nodes of the standard regular grid, thereby calculating the non-integer sampled coordinate grid after the flow field deformation. Since the non-integer coordinates typically fall between pixels in the original grid and cannot be directly read, a differentiable bilinear interpolation algorithm is used as the sampling strategy. The value of the sampling point is determined by calculating the weighted average of the feature values ​​of four adjacent integer pixels surrounding the target floating-point coordinates, where the weights are determined by the geometric distance between the sampling point and each adjacent pixel. For cases where the calculated sampling coordinates exceed the watershed grid boundary due to excessive flow velocity, a zero-filling strategy is used to handle the boundary-crossing area, indicating that the algal bloom material at that location has flowed out of the current computational domain or that no external material has flowed in, thus forming a dynamic sampling feature that accurately reflects the physical processes of fluid transport. This displacement guidance information essentially describes the possible migration direction and relative displacement trend of the water body and its carried cyanobacteria community within each spatial grid cell within the current time step. In practical applications, this displacement guidance information can reflect typical hydrodynamic phenomena such as downstream drift in the main channel and stagnation in the tributary backflow zone, thus providing a physical constraint basis for subsequent feature resampling.

[0055] After obtaining the displacement guidance information, the Lagrange translational flow convolution gate no longer uses fixed grid coordinates for convolution calculation. Instead, based on the displacement guidance information, it performs continuous coordinate offset mapping on the feature tensors representing spatial distribution in the spatiotemporal feature sequence of cyanobacterial blooms. This process is equivalent to tracking the features of cyanobacterial blooms in a Lagrange reference frame, causing the feature sampling points at each spatial location to shift with changes in the flow field, thus forming dynamic sampling features. In this way, the model can simulate the positional changes of cyanobacteria as they are transported along with the overall water body at the feature level, rather than relying solely on local static convolution receptive fields for approximation. After completing dynamic sampling, the dynamic sampling features are input into the convolution operation path to extract local spatial structure information under flow field constraints. The Lagrange translational flow convolution gate also introduces a gated fusion structure to weightedly integrate the dynamic sampling features with the current hidden state of the PredRNN. The gated fusion adaptively adjusts the contribution ratio between historical memory features and current flow-driven features, allowing the model to prioritize dynamically sampled features in regions with high flow velocities and rapid algal bloom migration, while retaining more historical hidden state information in regions with low flow velocities or stagnant areas. This avoids memory corruption caused by indiscriminate updates. Through this structural design, the Lagrange translational flow convolution gate ultimately outputs a value reflecting the spatial location migration of cyanobacterial blooms as they migrate with the flow field. This value serves as a crucial input for subsequent state updates of the PredRNN, enabling the model to continuously maintain awareness of spatial translation behavior throughout the time progression.

[0056] This embodiment effectively improves the problem of difficulty in explicitly characterizing the spatial migration driven by the flow field by constructing a Lagrange translational flow convolution gate with displacement guidance, dynamic sampling and gated fusion capabilities inside PredRNN. This allows the model to introduce spatial evolution constraints that conform to physical mechanisms while maintaining long-term memory capabilities, thereby improving the accuracy and stability of watershed cyanobacterial bloom evolution prediction under complex hydrodynamic conditions.

[0057] Furthermore, the specific implementation process for extracting distribution intensity fluctuations modulated by environmental factors, through a reaction-diffusion wave gate that includes environmental modulation, wave update, and diffusion constraints, includes:

[0058] The reaction-diffusion fluctuation gate jointly maps the feature channels representing chlorophyll concentration changes in the spatiotemporal feature sequence of algal blooms with environmental factors of the corresponding time slices to generate environmental modulation information corresponding to the feature channels of the hidden state. Based on the environmental modulation information, the feature components related to algal bloom intensity changes in the hidden state of PredRNN are updated with modulated temporal difference to form fluctuation intermediate features. The fluctuation intermediate features are subjected to diffusion constraint processing with temporal continuity, and the output is the distribution intensity fluctuation.

[0059] Specifically, the reaction-diffusion fluctuation gate is set in the hidden state update pathway of PredRNN. Its input includes feature channels in the spatiotemporal feature sequence of cyanobacterial blooms used to characterize changes in chlorophyll concentration, and environmental factor information corresponding one-to-one with the current time slice. The environmental factors include at least parameters such as water temperature and light intensity that have a direct regulatory effect on the evolution of cyanobacterial blooms. Their spatial distribution and temporal sampling have been aligned, enabling them to be jointly modeled with the bloom features in the same spatiotemporal coordinate system.

[0060] The reaction-diffusion wave gate first performs a joint mapping process, fusing and encoding the feature channels characterizing chlorophyll concentration changes with environmental factor information to generate environmental modulation information corresponding one-to-one with the feature channels of the PredRNN hidden state. This environmental modulation information is used to characterize the potential direction and intensity of the influence of environmental parameter changes on the evolution of cyanobacterial blooms under unchanged hydrodynamic conditions, such as the enhancing effect of evolution under high water temperature and high light conditions, or the attenuating effect of evolution under low temperature and low light conditions. After obtaining the environmental modulation information, the reaction-diffusion wave gate performs modulated temporal difference updates on the feature components in the PredRNN hidden state related to the change in bloom intensity based on this information. Specifically, the modulated temporal difference update adopts a first-order residual recursive calculation logic. First, the input environmental factors are processed through a hyperbolic tangent activation function, mapping them into an environmental modulation coefficient tensor with a numerical range between -1 and +1, to quantify the inhibitory or promoting effect of the environment on algal growth. Simultaneously, potential change features are extracted from the hidden state of the previous time step, and these features are multiplied element-wise with the environmental modulation coefficient tensor to obtain the environmentally weighted actual increment. This actual increment is then directly added to the hidden state value of the previous time step, completing the temporal differential update. The resulting feature is the fluctuation intermediate feature. The diffusion constraint processing aims to simulate the physical diffusion of biomass in water while ensuring temporal continuity. A depthwise separable convolution is used to spatially smooth the fluctuation intermediate feature. The convolution kernel is initialized with a weight distribution approximating the Laplace operator to simulate the diffusion equation. To prevent numerical abrupt changes, a forgetting gate mechanism similar to that in long short-term memory networks is introduced, generating a temporally smoothing gate with a value between zero and one. This gate is used to perform a weighted average of the spatially smoothed fluctuation feature and the original hidden state of the previous time step. Through this weighted fusion, it is enforced that the distribution intensity at the current time step cannot deviate from the historical state and produce non-physical drastic jumps, thus achieving the diffusion constraint of temporal continuity. This update process does not directly overwrite the original latent state, but rather characterizes the increment or decrease of chlorophyll concentration over time as environmental conditions change, thus forming a fluctuating intermediate feature that reflects the dynamic evolution of cyanobacterial blooms. In this way, the model can introduce intensity fluctuation information consistent with the evolution of blooms while maintaining the continuity of historical memory.

[0061] Considering that the evolution of cyanobacterial blooms typically exhibits both continuity and lag over time, the reaction-diffusion fluctuation gate applies a temporally continuous diffusion constraint to the intermediate fluctuation features it generates. This diffusion constraint suppresses unreasonable and drastic fluctuations within short timescales, ensuring that the cyanobacterial bloom evolution process exhibits a smooth and continuous evolutionary trend over time, thus better aligning with the objective laws governing the gradual accumulation or dissipation of actual cyanobacterial communities. The features processed by the diffusion constraint are then output as the distribution intensity fluctuation and participate in subsequent PredRNN state updates and prediction generation.

[0062] This embodiment introduces a reaction-diffusion wave gate, which includes environmental modulation, wave update, and diffusion constraints, into PredRNN to effectively decouple the cyanobacterial wave process from the spatial transport process. This enables the model to specifically model the intensity changes of algal blooms driven by environmental factors, thereby improving the problem that intensity wave characteristics are submerged in the overall spatiotemporal evolution and enhancing the performance of watershed cyanobacterial bloom evolution prediction in terms of intensity change characterization and long-term stability.

[0063] Furthermore, the specific implementation process of independently evolving and updating spatial location migration and distribution intensity fluctuations within an orthogonal subspace, and then converging them into a hybrid latent state of the watershed at the current moment through feature recoupling includes:

[0064] The spatial position migration output by the Lagrange translation flow convolution gate and the distribution intensity fluctuation output by the reaction diffusion wave gate are respectively mapped to mutually orthogonal feature subspaces. PredRNN state updates are performed along the time dimension in their respective subspaces, and independent temporal evolution is maintained. At the current time, the state representations in the two subspaces are aligned and fused by the feature recoupling operator to generate a hybrid hidden state that simultaneously contains spatial migration information and distribution intensity information.

[0065] Specifically, the spatial location migration is mainly constrained by the flow field velocity vector, and its temporal evolution has strong directional consistency and continuous translation characteristics; the distribution intensity fluctuation is mainly regulated by environmental factors, and its temporal evolution exhibits relatively slow but non-stationary intensity change characteristics.

[0066] To avoid mutual coupling and dilution of the two types of features during PredRNN state updates, the spatial location migration output from the Lagrange translation flow convolution gate is mapped to the first feature subspace, while the distribution intensity fluctuation output from the reaction-diffusion wave gate is mapped to the second feature subspace. The first and second feature subspaces are orthogonal in their feature dimensions. This orthogonal mapping ensures that the spatial location migration-related features and the intensity fluctuation-related features are independent at the representation level, structurally avoiding feature aliasing. Within their respective feature subspaces, PredRNN state updates are performed along the time dimension, ensuring that spatial location migration and distribution intensity fluctuation maintain independent temporal evolution in different state paths. The state update in the spatial location migration subspace focuses on capturing the continuous displacement process of the overall morphology of cyanobacterial blooms under the influence of the flow field, while the state update in the distribution intensity fluctuation subspace focuses on characterizing the intensity changes driven by the environment during the bloom evolution process. By maintaining two independent temporal evolution paths, the gradient interference and long-term memory degradation problems caused by unified hidden state updates are avoided. Specifically, the separate mapping operation is implemented through two 1x1 convolutional projection layers with non-shared parameters. These two projection layers map the spatial location migration and distribution intensity fluctuations of the input to a high-dimensional hidden layer space, respectively. To mathematically guarantee the orthogonality of the two subspaces, a decorrelation regularization term is introduced into the loss function during training. This regularization term calculates the cosine similarity matrix between the feature vectors of the two subspaces and forces its off-diagonal elements to approach zero, thereby constraining the feature basis of the two subspaces to remain linearly independent during optimization and preventing semantic confusion of features. In the independent evolution phase, the state updates within the two subspaces are executed by two sets of PredRNN recurrent units with independent weight parameters, without lateral connections between them, ensuring the decoupled propagation of spatial flow field information and biomass intensity information along the time axis. As for the alignment and fusion at the current time step, the feature recoupling operator adopts a gated fusion mechanism based on channel attention. The mechanism first concatenates the hidden states output from the two subspaces along the channel dimension to construct a joint feature tensor. Then, it uses global average pooling to compress the spatial dimension of the joint feature, generating a global context descriptor. The importance weight of each channel is calculated through a two-layer multilayer perceptron network. The joint feature tensor is adaptively weighted and recombined using this weight, and the weighted feature is compressed back to the dimension of the original hidden state through a 1x1 fusion convolutional layer. This process aligns and unifies the two independent evolution states while preserving key information.

[0067] At the current moment, the state representations in the two orthogonal subspaces are aligned and fused using a feature recoupling operator. This recoupling operator, while maintaining the integrity of the features of each subspace, converges spatial migration information and distribution intensity information into a unified hybrid latent state. This hybrid latent state can simultaneously reflect the spatial migration trend and intensity distribution fluctuations of cyanobacterial blooms. Through this recoupling process, the model neither loses spatial structural continuity nor weakens its sensitivity to changes in bloom intensity.

[0068] This embodiment effectively improves the information dissipation and prediction ambiguity caused by heterogeneous dynamic feature hybrid modeling by modeling spatial location migration and distribution intensity fluctuations separately in orthogonal subspaces and recoupling features at the current time node. This enables PredRNN to balance spatial migration accuracy and intensity change characterization in long-term algal bloom evolution forecasting tasks, thereby improving overall prediction performance and engineering applicability.

[0069] Furthermore, the specific implementation process of using a cross-step-length history recall mechanism to calculate the topological similarity between the mixed hidden state and each memory anchor point in the watershed history evolution database, and retrieving and aggregating historical evolution trend items that conform to chlorophyll concentration includes:

[0070] The cross-step historical recall mechanism inputs the current mixed latent state into the historical recall path and aligns it with the multi-timescale memory anchors stored in the watershed historical evolution database. Based on the joint representation of the mixed latent state in the dimensions of spatial structure distribution and algal bloom intensity change, it calculates the topological similarity between the mixed latent state and each memory anchor. The topological similarity results are used to filter historical evolution segments that match the current chlorophyll concentration, and the historical evolution segments are weighted and aggregated to output the historical evolution trend term.

[0071] Specifically, the cross-step historical recall mechanism is activated after obtaining the mixed latent state at the current moment. This mixed latent state is formed by the re-coupling of spatial location migration and distribution intensity fluctuations in the preceding steps, and can simultaneously characterize the comprehensive state of cyanobacterial blooms in both spatial structure distribution and intensity change dimensions. The mixed latent state is used as a query feature and input into the historical recall pathway for alignment and matching with a pre-constructed watershed historical evolution database.

[0072] The watershed historical evolution database was constructed during the model training phase by organizing and processing years of historical monitoring data from the target watershed. It internally stores memory anchors at multiple time scales. These memory anchors correspond to different typical algal bloom evolution processes, including short-term sudden expansion, medium-term stable maintenance, and long-term decay. During storage, they maintain a feature representation consistent with the spatial grid structure to ensure alignment with the current mixed latent state within the same feature space. Specifically, the multi-time scale division logic is based on different combinations of sliding window sampling intervals and coverage durations, aiming to cover different evolutionary patterns of cyanobacterial blooms, from hourly rapid aggregation and daily diurnal vertical migration to weekly seasonality. During construction, three time bases—short, medium, and long—were established, corresponding to high-frequency sampled short-term mutation sequences, diurnal fluctuation sequences containing complete diurnal cycles, and long-term trend sequences after downsampling, respectively, forming historical evolution segments of different lengths. For each of these historical evolution segments, the memory anchor is not a simple slice of the original data, but rather a high-dimensional mapping representation of that segment in the feature space. The specific construction method involves inputting the preceding subsequence of the extracted historical evolution fragment into a pre-trained feature extraction network, extracting the mixed hidden state tensor at the last time step as a memory anchor. This memory anchor is structurally represented as a three-dimensional feature tensor with a fixed number of channels and spatial resolution, possessing the same tensor dimension as the current mixed hidden state generated by the model during real-time inference, thus ensuring mathematical comparability in subsequent calculations. Simultaneously, a strict key-value pair indexing relationship is established between the memory anchor and the historical evolution fragment. The memory anchor serves as the "key," representing a specific precursor pattern of algal bloom evolution, used for retrieval and matching in the database, while the subsequent actual evolution data immediately following the anchor time in the historical evolution fragment serves as the "value." Through this structure, when the current mixed hidden state successfully matches a memory anchor in the database, the subsequent algal bloom evolution path that actually occurred under similar hydrodynamic and environmental precursor conditions in the past can be indexed through this key-value relationship, providing a trend reference for current predictions.

[0073] In the historical recall pathway, the cross-step historical recall mechanism first performs feature alignment processing on the current mixed latent state and each memory anchor point in the watershed historical evolution database, making them comparable in terms of spatial structure distribution and algal bloom intensity variation. Based on this alignment result, the topological similarity between the current mixed latent state and each memory anchor point is further calculated to measure the similarity between the current algal bloom evolution state and the historical typical evolution pattern in terms of spatial morphology and intensity variation trend. Through this topological similarity calculation process, the model can identify the evolution trajectory that is closest to the current state in historical data, rather than relying solely on local information from the most recent time step. Specifically, a first feature channel representing the algal bloom intensity distribution is extracted, and the Sobel operator is used to perform spatial convolution on this intensity distribution to extract a second feature channel representing the algal bloom boundary gradient and morphological skeleton. These two channels are then concatenated in the depth dimension to form a joint feature tensor that takes into account both numerical intensity and geometric structure. Based on this, the calculation of the topological similarity adopts a composite metric that includes a structural similarity index and Hausdorff distance. In the calculation process, firstly, for the first feature channel (intensity dimension), the statistical distance between the current state and the memory anchor point in terms of global mean and variance is calculated to quantify the difference in biomass intensity. Secondly, for the second feature channel (structure dimension), the gradient map is binarized using the Otsu method to extract the connected component contours of the algal bloom clusters. These two contours are regarded as a set of points in two-dimensional space, and the Hausdorff distance between them is calculated to measure the maximum mismatch between the two irregular geometric shapes in the topological space. After normalizing the intensity statistical distance and the structural Hausdorff distance, the topological similarity is generated.

[0074] After obtaining the topological similarity results, the cross-step historical recall mechanism further combines the current chlorophyll concentration level to filter the memory anchors in the watershed historical evolution database, retaining only historical evolution fragments that match the current state within the algal bloom intensity range. This avoids introducing historical patterns without practical reference value into the current prediction process. Multiple historical evolution fragments obtained from the filtering are then weighted and aggregated to generate a historical evolution trend term that reflects the possible evolution direction of the current state within a historical context.

[0075] This embodiment uses a cross-step-length historical recall mechanism to perform topological similarity matching between the current mixed hidden state and typical patterns in the watershed historical evolution database. This effectively compensates for the lack of information caused by PredRNN relying solely on short-term historical states for extrapolation. It enables the model to introduce historical priors across time scales during the prediction process, thereby improving the stability and trend consistency of watershed cyanobacterial bloom evolution forecasts in long-term time series prediction tasks.

[0076] Furthermore, by inputting historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed latent states, the specific implementation process for predicting the distribution field of algal bloom evolution in future periods includes:

[0077] The historical evolution trend term is jointly mapped with the current mixed hidden state to form an enhanced state representation, which is then input into the memory enhancement pathway. In the memory enhancement pathway, nonlocal information transfer across spatial regions is performed to supplement features in regions with sparse and insufficient information in the mixed hidden state, and the uncertainty in the state representation is compensated and adjusted based on the historical evolution trend term. The enhanced state representation is then input into the PredRNN output pathway to generate the algal bloom evolution distribution field corresponding to the future time period.

[0078] Specifically, after the retrieval and aggregation of historical evolution trend terms, these terms, as important prior information reflecting the potential evolution direction of the current algal bloom state within a historical context, are jointly mapped with the current hybrid latent state before being input into the memory enhancement pathway. This joint mapping process aligns and fuses the hybrid latent state and historical evolution trend terms within a unified feature space, forming an enhanced state representation. This enhanced state representation retains the current observation-driven features while incorporating historical evolution information across time scales.

[0079] After receiving the enhanced state representation, the memory enhancement pathway performs nonlocal information transfer across spatial regions. Specifically, information-sparse and insufficient regions are defined as grid positions in the current mixed latent state feature map where the activation response value is lower than the global average threshold. These positions typically correspond to areas where algal blooms have not yet clearly erupted but have potential evolutionary risks. To supplement features in these regions, three independent convolutional layers are used to map the input features into query vectors, key vectors, and value vectors, respectively. The query vector originates from the current mixed latent state, while the key and value vectors originate from the retrieved historical evolutionary trend terms. The dot product similarity between the query vector and all key vectors in the global space is calculated to generate a global affinity matrix, which quantifies the association strength between each current grid point and any spatial position in the historical evolutionary trend. The affinity matrix is ​​normalized using the Softmax function to filter out highly correlated feature responses, thereby suppressing interference from distant irrelevant noise. Based on the normalized affinity matrix, the value vectors of the historical evolutionary trend are aggregated, and the aggregated features are superimposed back into the original mixed latent state as residual terms. In this way, regions with sparse information at the present moment can borrow feature information from high-response regions with similar environmental backgrounds or evolution patterns in the past, thereby completing non-local feature enhancement and filling without disrupting local continuity.

[0080] In real-world waters, cyanobacterial blooms often exhibit co-evolutionary characteristics in discontinuous spatial locations such as main channels, backflow zones, or bays. The nonlocal information transfer mechanism effectively captures these correlations, compensating for information gaps caused by insufficient observation of single points or local areas. While performing nonlocal information transfer, the memory enhancement pathway compensates for uncertainties in the enhanced state representation based on historical evolution trend terms. These uncertainties primarily stem from the gradual accumulation of errors within the model's internal states during long-term extrapolation and the lack of information in sparse regions of the observed data. By introducing historical evolution trends highly similar to the current state, the prediction direction and intensity are constrained, preventing unreasonable sharp shifts or over-smoothing in subsequent inference processes.

[0081] After completing the nonlocal information supplementation and uncertainty compensation, the enhanced state representation is input into the output path of the PredRNN to generate the algal bloom evolution distribution field corresponding to the future time period. The algal bloom evolution distribution field can maintain a relatively clear structural boundary in space and can continue a reasonable evolution trend in time, thus meeting the dual requirements of accuracy and stability for watershed-scale algal bloom forecasting.

[0082] This embodiment introduces historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed hidden states. This effectively improves the information decay and result ambiguity problems that PredRNN is prone to in long-term algal bloom evolution prediction. It enables the prediction model to output a stable and reliable algal bloom evolution distribution field in complex watershed environments, thereby improving the overall prediction performance and engineering application value.

[0083] Example 2:

[0084] This embodiment provides a watershed cyanobacterial bloom evolution prediction system based on an improved PredRNN, referring to... Figure 2 The system includes a data processing module, a feature extraction module, a history recall module, and an evolution prediction module.

[0085] The data processing module acquires monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performs spatiotemporal alignment and gridded interpolation, and constructs a spatiotemporal feature sequence of algal blooms, which is then input into PredRNN.

[0086] In the feature extraction module of PredRNN, spatial position migration constrained by flow field velocity vector is extracted through a Lagrange translational flow convolution gate that includes displacement guidance, dynamic sampling and gated fusion; and distribution intensity fluctuations that are regulated by environmental factors are extracted through a reaction-diffusion fluctuation gate that includes environmental modulation, fluctuation update and diffusion constraint.

[0087] The historical recall module independently evolves and updates the spatial location migration and distribution intensity fluctuations within the orthogonal subspace, and converges them into a mixed hidden state of the watershed at the current moment through feature recoupling; using the cross-step historical recall mechanism, it calculates the topological similarity between the mixed hidden state and each memory anchor point in the watershed historical evolution database, and retrieves and aggregates historical evolution trend items that conform to chlorophyll concentration.

[0088] The evolution prediction module inputs historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed latent state, and predicts the distribution field of algal bloom evolution in future periods.

[0089] Furthermore, the data processing module is configured to receive continuous observation data from multiple types of monitoring devices within the target watershed. This observation data includes at least chlorophyll concentration data to characterize the intensity of cyanobacterial blooms, flow velocity vector data to describe water transport characteristics, and environmental factor data to reflect the environmental conditions for cyanobacterial growth. The module performs integrity and validity checks on all types of raw monitoring data to identify missing data segments and abnormal fluctuations caused by sensor anomalies, communication interruptions, or external interference, thereby reducing the interference of noise data on subsequent model training and inference processes. After completing the basic verification, the data processing module performs quality control processing on the raw monitoring data. By comparing historical statistical distributions, it removes outliers that significantly deviate from the normal range. For data gaps occurring within short timescales, it performs reasonable compensation by combining valid observations within the temporal or spatial neighborhood, thus forming a continuous and stable raw observation dataset. Based on a unified time reference, the module performs frequency resampling and time-series alignment processing on all types of observation data, ensuring that chlorophyll concentration, flow velocity vector, and environmental factors have consistent timestamp representations within the same time slice. This time alignment operation ensures that the correspondence between the features of each channel in the time dimension is clear during the subsequent PredRNN modeling process, thus avoiding the introduction of implicit time bias by multi-source heterogeneous data.

[0090] In the spatial dimension, the data processing module constructs a regularized spatial grid structure covering the target watershed and maps discretely distributed monitoring section data into this spatial grid. Through spatial interpolation, the originally irregularly distributed point observation data is transformed into a continuous spatial raster representation, thus meeting the PredRNN's requirement for a regular spatial input structure. The generated spatial raster data is dimensionless to eliminate the influence of differences in numerical scales among different physical quantities. The dimensionless chlorophyll concentration, flow velocity vector components, and environmental factors are stacked along the channel dimension to form a multidimensional state feature tensor, used to comprehensively characterize the state of cyanobacterial blooms, hydrodynamic conditions, and environmental driving factors. This multidimensional state feature tensor maintains the integrity of various information types while possessing a unified data structure, suitable for input into the subsequent PredRNN network for spatiotemporal modeling. In terms of temporal organization, the data processing module performs sliding window truncation on the multidimensional state feature tensor at consecutive moments according to the natural temporal order of bloom evolution, constructing a spatiotemporal feature sequence of blooms. The spatiotemporal feature sequence of algal blooms is used as input to PredRNN, enabling the model to learn both spatial distribution changes and temporal evolution patterns during the state recursion process.

[0091] Furthermore, the feature extraction module receives the spatiotemporal feature sequence of algal blooms output by the data processing module. This sequence consists of multidimensional state feature tensors across consecutive time slices, where each feature channel corresponds to chlorophyll concentration, flow velocity vector components, and environmental factors, respectively. The module first performs channel consistency checks and dimension rearrangement on the multidimensional state feature tensors to ensure stable spatial alignment of different physical quantities during subsequent feature transformations. During spatial feature extraction, the module performs local spatial receptive field modeling on the state feature tensors within each time slice. By encoding the correlation between adjacent spatial grid cells, the model can identify the spatial continuity, boundary changes, and local high-value area structure of algal bloom distribution. This spatial feature extraction not only focuses on single-scale spatial morphology but also implicitly preserves multi-scale spatial responses in the feature channels, allowing the overall morphology and fine-grained structure of algal bloom clusters to be characterized simultaneously. This avoids the problem in existing methods of only capturing coarse-scale distributions while ignoring local anomalies. In terms of temporal feature extraction, the module performs correlation modeling of spatial features at continuous time points along the time dimension, enabling the trends in algal bloom concentration, migration velocity, and environmental driving responses to be embedded in the feature representation in a temporally continuous form. Through this temporal feature encoding method, the model can distinguish between short-term fluctuations and continuous evolutionary trends, avoiding prediction results that are lagging or overly smoothed due to insufficient temporal modeling capabilities. Cross-channel correlation modeling is performed on feature channels with different physical meanings, allowing chlorophyll concentration variation features to form an effective coupling relationship with flow field velocity vectors and environmental factor features. Through this cross-channel feature extraction, the model can learn the intrinsic correlation between algal bloom evolution and hydrodynamic conditions and environmental changes.

[0092] Furthermore, the historical retrieval module receives the current hybrid latent state, which comprehensively represents both the spatial migration and intensity fluctuations of cyanobacterial blooms, reflecting the overall state of bloom distribution in the current watershed. The module internally constructs a watershed historical evolution database to store multi-timescale memory anchors extracted from historical monitoring periods. These memory anchors are generated from historical bloom spatiotemporal feature sequences using a feature mapping method consistent with the current model structure, ensuring the comparability of historical information and the current state in the feature space. During historical retrieval, the module first inputs the current hybrid latent state into the historical retrieval path and performs feature alignment processing with each memory anchor in the watershed historical evolution database. This feature alignment process eliminates feature scale offsets caused by observation conditions, seasonal variations, or differences in monitoring density, enabling similarity assessment between the current and historical states within a unified feature representation space. After feature alignment, the module calculates the topological similarity between the hybrid latent state and each memory anchor based on the joint representation of the hybrid latent state in the dimensions of spatial structure distribution and bloom intensity variation. The topological similarity is used to characterize the degree of similarity between the current algal bloom distribution pattern and its historical evolution pattern in terms of spatial layout, expansion direction, and intensity trends. This avoids the problem of ignoring the overall structural relationship by matching only numerical proximity. Based on the topological similarity calculation results, the module selects historical evolution segments from the watershed's historical evolution database that match the current chlorophyll concentration level and its changing trend. These historical evolution segments cover evolution processes at different time scales, including algal bloom processes that expand or decline rapidly in the short term, as well as seasonal evolution trends across cycles. By weighted aggregation of the selected historical evolution segments, the historical recall module generates historical evolution trend items that reflect the current direction of algal bloom development.

[0093] Furthermore, the evolution prediction module receives an enhanced state representation output by the memory enhancement pathway. This enhanced state representation, based on the original mixed hidden states, incorporates spatial location migration information, distribution intensity fluctuation information, and historical evolution trend information, comprehensively depicting the driving factors and development trends of current watershed algal bloom evolution. The module first performs a temporal consistency check and feature reorganization on the enhanced state representation to ensure the continuity and stability of the hidden state evolution during multi-step prediction. During prediction execution, the module recursively applies the PredRNN state update mechanism along the time dimension, allowing the enhanced state representation to continuously evolve over multiple prediction steps. Combining the historical prior information and nonlocal compensation information contained in the enhanced state, the model maintains a global awareness of algal bloom evolution trends during temporal extrapolation, thus preventing the prediction results from gradually deviating from the actual evolution path over long periods. Spatially, the module generates an algal bloom evolution distribution field consistent with the watershed spatial grid based on the spatial structure information retained in the enhanced state representation. The algal bloom evolution distribution field can reflect the spatial location, diffusion direction and concentration changes of cyanobacterial blooms within the watershed at each prediction time slice, giving the prediction results good spatial interpretability.

[0094] This embodiment effectively addresses the problem that raw data cannot directly support deep spatiotemporal prediction models by performing quality control, spatiotemporal alignment, structured mapping, and feature organization on multi-source heterogeneous watershed monitoring data. This lays a solid data foundation for improving the high-precision algal bloom evolution forecasting of PredRNN. By performing multi-level and multi-dimensional feature modeling on the spatiotemporal feature sequences of algal blooms, the insufficient expressive power of the original features is effectively improved, enabling PredRNN to perform state evolution and prediction in a more discriminative feature space. By introducing historical evolution information across time scales during the prediction process and performing precise matching and aggregation based on topological similarity, the problem of insufficient utilization of historical information due to relying solely on short-term time-series modeling is effectively improved, giving PredRNN stronger long-term memory and trend preservation capabilities in algal bloom evolution prediction. By performing stable recursion and spatially consistent output on the enhanced state representation, the problems of accuracy degradation and uncertainty amplification in algal bloom evolution prediction during long-term extrapolation are effectively improved, thereby enhancing the engineering value and decision support capability of watershed cyanobacterial bloom evolution forecasting in practical applications.

[0095] Example 3:

[0096] This embodiment deploys the above-mentioned improved PredRNN-based watershed cyanobacterial bloom evolution prediction method and system at a watershed cyanobacterial bloom evolution prediction center, referring to... Figure 3 This enables the prediction of the evolution of cyanobacterial blooms in the watershed.

[0097] Furthermore, multi-source monitoring data for the target watershed were collected and processed. This multi-source monitoring data included chlorophyll concentration monitoring data acquired by a fluorescence sensor, two-dimensional flow field velocity vector data acquired by an acoustic Doppler current profiler, and environmental factor data such as water temperature, light intensity, and nutrients acquired by meteorological and water quality sensors. In actual data collection, the time resolution was one hour for chlorophyll concentration, two hours for flow velocity vectors, and half an hour for environmental factors. Abrupt noise was eliminated through quality control and anomaly removal operations, and missing data was appropriately supplemented. Frequency resampling and time synchronization were performed based on a unified time reference, ensuring consistent temporal representation of various monitoring data within the same time slice. A regularized spatial grid covering the entire watershed was constructed, and the discretely distributed monitoring data was mapped to each grid node through spatial interpolation, generating continuous spatial fields for chlorophyll concentration, flow field velocity vectors, and environmental factors. After dimensionless processing, various spatial fields are stacked along the feature channel dimension to form a multidimensional state feature tensor, which is then organized into a spatiotemporal feature sequence of algal bloom according to the time evolution order.

[0098] Furthermore, spatial, temporal, and cross-channel joint modeling is performed on the spatiotemporal feature sequences of algal blooms, enabling chlorophyll concentration variation characteristics to form a highly correlated expression with flow field transport characteristics and environmental driving characteristics. Through this feature extraction process, the model can explicitly distinguish between spatial migration phenomena caused by algal bloom diffusion and intensity distribution phenomena caused by environmental changes, thus laying the foundation for subsequent state decoupling and evolution. In the main structure of PredRNN, this embodiment introduces a Lagrange translational flow convolution gate and a reaction-diffusion wave gate. The Lagrange translational flow convolution gate jointly maps the spatiotemporal feature sequences of algal blooms with the flow field velocity vectors of the corresponding time slices, generating displacement guidance information consistent with the spatial grid coordinates, and dynamically shifting the sampling position of the feature tensor based on this displacement guidance information, thereby extracting the distribution characteristics of algal blooms that spatially migrate with water flow. Through this mechanism, the model can accurately capture the true migration path of cyanobacterial masses under conditions of significant flow velocity changes.

[0099] Furthermore, the reaction-diffusion fluctuation gate jointly modulates the feature channels characterizing chlorophyll concentration changes with environmental factors, performing environmentally constrained temporal updates on the feature components related to bloom intensity changes in the PredRNN hidden state. During the update process, diffusion continuity constraints are introduced to ensure the hidden state accurately reflects the bloom evolution process of cyanobacteria under different environmental conditions. In the state fusion stage, the spatial position migration output by the Lagrange translation flow convolution gate and the distribution intensity fluctuation output by the reaction-diffusion fluctuation gate are mapped to mutually orthogonal feature subspaces, and PredRNN state evolution is performed independently within each subspace. By introducing a feature recoupling operator at the current time, the state representations in the two subspaces are aligned and fused, generating a hybrid hidden state that simultaneously contains spatial migration and intensity distribution information, thereby avoiding modeling biases caused by interference from different physical mechanisms.

[0100] Furthermore, the current mixed latent state is input into the historical retrieval pathway, and feature alignment and topological similarity calculation are performed with multi-timescale memory anchors stored in the watershed historical evolution database. The watershed historical evolution database is constructed from monitoring data of cyanobacterial blooms in the watershed over the past five years, covering multiple high-incidence and low-incidence evolution scenarios. Through topological similarity matching, the model can retrieve historical evolution fragments that match the current chlorophyll concentration level and spatial distribution pattern, and perform weighted aggregation to form a historical evolution trend term. The memory enhancement pathway jointly maps the historical evolution trend term with the current mixed latent state, performing non-local information supplementation and trend correction on areas with sparse information or high uncertainty in the mixed latent state, thereby generating an enhanced state representation. This enhanced state representation is input into the evolution prediction module, which generates the bloom evolution distribution field for multiple future time periods through the recursive output pathway of PredRNN.

[0101] This embodiment improves the problems of insufficient characterization of migration mechanisms, inadequate response to intensity distribution, and decreased accuracy in medium- and long-term predictions in the prediction of cyanobacterial bloom evolution by introducing multiple mechanisms such as flow field constraints, environmental modulation, and historical recall. It enhances the accuracy, stability, and engineering application value of watershed cyanobacterial bloom evolution forecasts in complex real-world scenarios.

[0102] It should be clarified that the embodiments described above are merely exemplary and are intended to aid in understanding the present invention, not to limit it. Those skilled in the art can make various changes and modifications after grasping the core ideas of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A watershed cyanobacterial bloom evolution prediction method based on improved PredRNN, characterized in that, include: The monitoring data of chlorophyll concentration, flow field velocity vector and environmental factors in the target watershed are acquired and spatiotemporally aligned and interpolated in a gridded manner. The spatiotemporal feature sequence of algal bloom is constructed and input into PredRNN. In PredRNN, the specific implementation process for extracting spatial position migration constrained by flow field velocity vectors through a Lagrange translational flow convolution gate that includes displacement guidance, dynamic sampling, and gated fusion includes: the Lagrange translational flow convolution gate jointly maps the input spatiotemporal feature sequence of algal bloom with the flow field velocity vector of the corresponding time slice to generate displacement guidance information consistent with the spatial grid coordinates; based on the displacement guidance information, the sampling coordinates of the feature tensors representing spatial distribution in the spatiotemporal feature sequence of algal bloom are continuously offset and mapped to form dynamic sampling features that change with the flow field; the dynamic sampling features are input into the convolutional operation path, and the dynamic sampling features are added to the current hidden state of PredRNN through gated fusion. The weighted integration outputs spatial location migration of cyanobacteria as they migrate with the flow field. The specific implementation process of extracting distribution intensity fluctuations regulated by environmental factors through a reaction-diffusion wave gate, which includes environmental modulation, wave update, and diffusion constraints, includes: the reaction-diffusion wave gate jointly maps the feature channels characterizing chlorophyll concentration changes in the spatiotemporal feature sequence of the algal bloom with the environmental factors of the corresponding time slices, generating environmental modulation information corresponding to the latent state feature channels; based on the environmental modulation information, performing modulated temporal differential updates on the feature components related to algal bloom intensity changes in the latent state of the PredRNN, forming intermediate wave features; and applying temporally continuous diffusion constraints to the intermediate wave features, outputting the distribution intensity fluctuations. The specific implementation process of independently evolving and updating spatial location migration and distribution intensity fluctuations within orthogonal subspaces, and then converging them into a hybrid latent state of the watershed at the current moment through feature recoupling includes: mapping the spatial location migration output by the Lagrange translation flow convolution gate and the distribution intensity fluctuation output by the reaction-diffusion fluctuation gate to mutually orthogonal feature subspaces respectively; performing PredRNN state updates along the time dimension in their respective subspaces and maintaining independent temporal evolution; at the current moment, aligning and fusing the state representations in the two subspaces through the feature recoupling operator to generate a hybrid latent state that simultaneously contains spatial migration information and distribution intensity information; and using a cross-step-length history recall mechanism to calculate the topological similarity between the hybrid latent state and each memory anchor point in the watershed historical evolution database, and retrieving and aggregating historical evolution trend items that conform to chlorophyll concentration. By inputting historical evolution trend terms into the memory enhancement pathway, nonlocal information supplementation and uncertainty compensation are provided for the mixed latent states, and the distribution field of algal bloom evolution in future periods is predicted.

2. The method for predicting the evolution of cyanobacterial blooms in watersheds based on improved PredRNN according to claim 1, characterized in that, The specific implementation process of acquiring monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performing spatiotemporal alignment and gridded interpolation, and constructing a spatiotemporal feature sequence of algal blooms as input to PredRNN includes: A multi-source heterogeneous sensor array was used to collect chlorophyll fluorescence intensity values ​​at monitoring sections, flow velocity vector components measured by acoustic Doppler current profilers, and environmental parameters such as light and temperature. Outliers were removed using statistical filtering algorithms, and missing data were filled using neighborhood mean to form the original observation dataset. Based on a unified time reference, the original observation dataset was resampled at frequency and synchronized with temporal phase to eliminate time delays between multi-source heterogeneous data. A regularized Euclidean grid covering the entire watershed was constructed, and a spatial interpolation algorithm was used to map the original observation datasets of discrete monitoring stations to the corresponding grid nodes, generating a continuous spatial raster containing chlorophyll concentration, flow velocity vector, and environmental factors. The continuous spatial raster was subjected to max-min dimensionless normalization, and chlorophyll concentration, flow velocity vector, and environmental factors were stacked along the channel dimension to construct a multidimensional state feature tensor. The multidimensional state feature tensor was extracted in continuous time sequence using a sliding window according to the temporal evolution order and combined to form a spatiotemporal feature sequence of algal bloom.

3. The method for predicting the evolution of cyanobacterial blooms in watersheds based on improved PredRNN according to claim 1, characterized in that, The specific implementation process of using a cross-step-length history recall mechanism to calculate the topological similarity between the mixed hidden state and each memory anchor point in the watershed history evolution database, and to retrieve and aggregate historical evolution trend items that conform to chlorophyll concentration includes: The cross-step historical recall mechanism inputs the current mixed latent state into the historical recall path and aligns it with the multi-timescale memory anchors stored in the watershed historical evolution database. Based on the joint representation of the mixed latent state in the dimensions of spatial structure distribution and algal bloom intensity change, it calculates the topological similarity between the mixed latent state and each memory anchor. The topological similarity results are used to filter historical evolution segments that match the current chlorophyll concentration, and the historical evolution segments are weighted and aggregated to output the historical evolution trend term.

4. The method for predicting the evolution of cyanobacterial blooms in watersheds based on improved PredRNN according to claim 1, characterized in that, The specific process of inputting historical evolution trend terms into the memory enhancement pathway to supplement mixed latent states with nonlocal information and compensate for uncertainty, and predicting the distribution field of future algal bloom evolution includes: The historical evolution trend term is jointly mapped with the current mixed hidden state to form an enhanced state representation, which is then input into the memory enhancement pathway. In the memory enhancement pathway, nonlocal information transfer across spatial regions is performed to supplement features in regions with sparse and insufficient information in the mixed hidden state, and the uncertainty in the state representation is compensated and adjusted based on the historical evolution trend term. The enhanced state representation is then input into the PredRNN output pathway to generate the algal bloom evolution distribution field corresponding to the future time period.

5. A watershed cyanobacterial bloom evolution prediction system based on improved PredRNN, wherein the system is applied to the watershed cyanobacterial bloom evolution prediction method based on improved PredRNN as described in any one of claims 1 to 4, characterized in that, include: The data processing module acquires monitoring data on chlorophyll concentration, flow field velocity vector, and environmental factors in the target watershed, performs spatiotemporal alignment and gridded interpolation, and constructs a spatiotemporal feature sequence of algal blooms, which is then input into PredRNN. In the feature extraction module of PredRNN, spatial position migration constrained by flow field velocity vector is extracted through a Lagrange translational flow convolution gate that includes displacement guidance, dynamic sampling and gated fusion; and distribution intensity fluctuations that are regulated by environmental factors are extracted through a reaction-diffusion fluctuation gate that includes environmental modulation, fluctuation update and diffusion constraint. The historical recall module independently evolves and updates the spatial location migration and distribution intensity fluctuations within the orthogonal subspace, and converges them into a mixed hidden state of the watershed at the current moment through feature recoupling; using the cross-step historical recall mechanism, it calculates the topological similarity between the mixed hidden state and each memory anchor point in the watershed historical evolution database, and retrieves and aggregates historical evolution trend items that conform to chlorophyll concentration. The evolution prediction module inputs historical evolution trend terms into the memory enhancement pathway to supplement nonlocal information and compensate for uncertainty in the mixed latent state, and predicts the distribution field of algal bloom evolution in future periods.

Images