N-BEATSx plankton prediction method fused with semantic model

Abstract

The invention relates to an N-BEATSx plankton prediction method fusing a semantic model. The method comprises the following steps: acquiring an observation data set in a current environment; performing multi-scale interpolation processing on the observation data set according to interpolation strategies corresponding to different interval durations to obtain a target observation data set comprising target multi-source environment data and target plankton data; and in the improved N-BEATSx network fused with the semantic model, performing cross attention operation and residual operation on the target multi-source environment data and the target plankton data to obtain a prediction result based on the plankton variation in the current environment. By adopting the method, semantic fusion and dynamic prediction of the environmental data and the biological data can be realized.

Description

Technical Field

[0001] This application relates to the field of ecological environment monitoring, and in particular to an N-BEATSx planktonic prediction method that incorporates a semantic model. Background Technology

[0002] In the field of ecological and environmental monitoring, time series analysis based on environmental parameters and plankton data is used to predict the changing trends of plankton.

[0003] In related data prediction methods, trend estimation is achieved by statistical fitting of historical observation data or by traditional neural network calculations. However, such methods rely only on numerical feature associations and lack an understanding of the semantic relationships of the data. As a result, the models have difficulty accurately capturing the deep dependency structure between the environment and plankton, affecting the accuracy and stability of the prediction results. Summary of the Invention

[0004] Therefore, it is necessary to provide an N-BEATSx planktonic prediction method, system, computer device, and computer-readable storage medium that integrates semantic models to address the aforementioned technical problems.

[0005] Firstly, this application provides an N-BEATSx planktonic prediction method that integrates semantic models, including: Obtain the observation dataset under the current environment, which includes multi-source environmental data and plankton data; Based on the interpolation strategies corresponding to different interval durations, the observation dataset is subjected to multi-scale interpolation processing to obtain the target observation dataset, which includes target multi-source environmental data and target plankton data; In the improved N-BEATSx network that pre-integrates semantic models, cross-attention and residual operations are performed on the target multi-source environmental data and the target plankton data to obtain the prediction results based on the change in plankton in the current environment.

[0006] Furthermore, in the improved N-BEATSx network of the fusion semantic model involved in the above method: on the one hand, the fully connected network (FCN), temporal convolutional network (TCN), and Hadamard product operation in the BLOCK block of the conventional N-BEATSx network are removed, and a cross-attention mechanism and a feed-forward neural network (FFN) are introduced in the BLOCK block; on the other hand, the forward prediction and backward prediction in the BLOCK block of the conventional N-BEATSx network are replaced by predictive linear layers and regression linear layers, respectively.

[0007] Based on the improved structure in the improved N-BEATSx network of the above-mentioned fusion semantic model, cross-attention operation and residual operation are performed to realize the semantic fusion and dynamic prediction of environmental data and biological data, making the analysis of plankton change trends more accurate, coherent and with stable explanatory logic.

[0008] Secondly, this application also provides an N-BEATSx planktonic prediction system that integrates semantic models, comprising: The acquisition module is used to acquire the observation dataset under the current environment, which includes multi-source environmental data and plankton data; The interpolation module is used to perform multi-scale interpolation processing on the observation dataset according to the interpolation strategies corresponding to different interval durations, so as to obtain the target observation dataset, which includes target multi-source environmental data and target plankton data. The prediction module is used to perform cross-attention and residual operations on the target multi-source environmental data and the target plankton data in an improved N-BEATSx network that pre-integrates semantic models, to obtain the prediction result based on the change in plankton in the current environment.

[0009] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the above steps.

[0010] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the above steps.

[0011] The aforementioned N-BEATSx plankton prediction method, system, computer equipment, and computer-readable storage medium based on the fusion semantic model first establishes an observation dataset by collecting and synchronizing multi-source environmental data and plankton data, thus ensuring the temporal consistency and integrity of the data foundation for subsequent modeling. Next, multi-scale interpolation processing is performed on the observation dataset according to interpolation strategies corresponding to different time intervals to obtain a target observation dataset containing target multi-source environmental data and target plankton data, thereby achieving data continuity and temporal scale alignment, providing a unified input for feature analysis. Furthermore, the improved N-BEATSx network based on the fusion semantic model performs cross-attention and residual operations on the two types of data, allowing the impact of environmental changes on the biological change rate sequence of plankton to be expressed in a unified structure, and generating a plankton change prediction result based on future time periods. Based on this, the entire technical solution achieves semantic fusion and dynamic prediction of environmental and biological data, making the analysis of plankton change trends more accurate, coherent, and possessing a stable interpretive logic. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the N-BEATSx planktonic prediction method that incorporates a semantic model in one embodiment. Figure 2 This is a schematic diagram of the structure of a dual residual block in an improved N-BEATSx network that integrates semantic models in one embodiment; Figure 3 This is a block diagram of the N-BEATSx planktonic prediction system that incorporates a semantic model in one embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0015] In one exemplary embodiment, such as Figure 1As shown, an N-BEATSx planktonic prediction method based on a fused semantic model is provided. This embodiment illustrates the method by applying it to a server. It is understood that this method can also be applied to a terminal, or to a system including both a terminal and a server, and can be implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps S101 to S104.

[0016] Step S101: Obtain the observation dataset under the current environment. The observation dataset includes multi-source environmental data and plankton data.

[0017] The current environment refers to the specific aquatic environment for which phytoplankton abundance prediction is to be performed, such as the aquatic range of a lake, river, or ocean area in a specific season; the observation dataset in the current environment refers to the raw data set obtained by synchronously collecting and summarizing environmental monitoring equipment and phytoplankton sampling equipment within a specific time period and spatial range.

[0018] Among them, multi-source environmental data refers to different types of environmental observation data obtained under the current environment, such as various environmental parameters such as water temperature, salinity, dissolved oxygen content, pH, turbidity, and chlorophyll content.

[0019] Among them, plankton data represents the observational data on the quantity of different types of plankton in the current environment, such as copepods, echinoderms, Nostoc, jellyfish, Noctiluca scintillans, shrimp, Phaeocystis, arrowworms, and tunicates.

[0020] Step S102: According to the interpolation strategies corresponding to different interval durations, multi-scale interpolation processing is performed on the observation dataset to obtain the target observation dataset, which includes target multi-source environmental data and target plankton data.

[0021] The interpolation strategies corresponding to different interval lengths represent a set of strategies for selecting numerical reconstruction methods that match the interval length to address the differences in sampling intervals in each time series of the observation dataset. These strategies are used to fill in missing values ​​in the time dimension, maintain data continuity, and ensure that data at different time resolutions can be recovered in a reasonable manner.

[0022] For example, for the observation dataset obtained in the aforementioned steps, multi-scale interpolation processing is performed based on the sampling characteristics of the data in the observation dataset at different intervals, using an interpolation strategy corresponding to the interval duration. That is, since there are significant differences in response speed between different types of multi-source environmental data and plankton data, a uniform interpolation strategy would cause trend shifts; therefore, corresponding interpolation strategies need to be matched for different interval durations. Specifically, throughout the process, different interpolation strategies are determined based on the matching relationship between the time index and the sampling interval to transform the original discrete data points into a continuous sequence. For example, a local trend fitting method is used for short-interval data, and an overall trend extension method is used for long-interval data, so that the data forms a logically continuous sequence structure in different time dimensions. Thus, multi-scale interpolation processing of multi-source environmental data yields target multi-source environmental data, and multi-scale interpolation processing of plankton data yields target plankton data. Furthermore, a target observation dataset with a complete temporal structure at multiple time scales is generated based on the target multi-source environmental data and the target plankton data.

[0023] Based on this, the target multi-source environmental data and target plankton data not only retain the numerical characteristics of the original observations, but also ensure that the changing trends over different time periods are traceable and comparable at the computational level.

[0024] Step S103: In the improved N-BEATSx network that pre-integrates semantic models, cross-attention operation and residual operation are performed on the target multi-source environmental data and the target plankton data to obtain the prediction results based on the change in plankton in the current environment.

[0025] Among them, the improved N-BEATSx network with fusion semantic model introduces a semantic association modeling structure on the basis of the original N-BEATSx time series decomposition framework. This structure is used to establish a semantic relationship chain between environmental factors and plankton changes that can be explicitly used by the network. For example, by adding a cross-attention structure to each double residual block, the semantic dependence of a certain environmental factor on which type of plankton changes is more influential at what time period is expressed. This makes the prediction process based not only on numerical patterns but also on the semantic relationship between environment and organism.

[0026] For example, the target multi-source environmental data and target plankton data obtained in the aforementioned steps are simultaneously input into an improved N-BEATSx network that fuses semantic models, and correlation calculations are performed under a unified time series structure. Specifically, firstly, the temporal distribution and numerical structure of the two types of input sequences are organized so that the environmental dimension and the plankton dimension can be identified and referenced within the same sequence framework. Subsequently, the network establishes semantic connections between the two types of data based on cross-attention operations, enabling the plankton biodiversity change rate sequence to actively focus on key parts of the environmental data related to its change trend during the calculation process; wherein, the biodiversity change rate sequence represents the serialized expression of the magnitude of plankton quantity change in adjacent time periods in the target plankton data; wherein, the cross-attention operation here plays the role of information alignment, allowing environmental data from different time locations to participate in the interpretation of historical change structures in a weighted manner, thereby constructing combined features for future prediction.

[0027] Next, after the cross-attention operation, the network preserves the basic trend structure of the input sequence through residual operation, so that the network can maintain the continuity and overall shape of the bio-change rate sequence of plankton when learning changes caused by environmental factors. The reason for the design of residual operation is that the change in plankton is often composed of long-term trends and short-term fluctuations. If the original structure is not preserved, the prediction process will deviate from the actual change trajectory. By combining the newly generated features with the original bio-change rate sequence in a residual manner, the network can stably superimpose environmental influences in structure without destroying the inherent law of the bio-change rate sequence.

[0028] Thus, as the multi-layered structure sequentially completes cross-attention and residual operations, the network gradually forms a composite expression that integrates environmental factors and historical changes in biomass, enabling the output to simultaneously reflect both the influence of environmental drivers and the continuity of the time series itself. Ultimately, predictions characterizing future phytoplankton changes are obtained from the network's endpoints, serving as a quantitative basis for subsequent ecological analysis. Based on this, the entire process combines semantic fusion mechanisms with residual structure preservation, ensuring that the prediction process possesses a clear source of association and a stable sequence expression.

[0029] Therefore, the improved N-BEATSx network that integrates semantic models has at least the following three advantages: (1) Direct and efficient feature fusion: By establishing a dynamic connection between environmental data and biological change rate sequence in the same computational path through cross-sequence interaction structure, the indirect method of traditional independent processing and recombination is avoided, so that the influence of each data source in the prediction can be adaptively adjusted, and more compact, direct and efficient feature fusion can be achieved. (2) Stronger representation ability: Through the global information citation method, the network can capture the cross-time period and non-linear dependence between environmental factors and the historical changes in the number of organisms in a unified structure, making the internal feature expression more complete and enhancing the modeling depth of multi-factor coupling relationship, which is conducive to improving the ability to express future changes. (3) Simplified architecture: The original multi-path structure is integrated into a single interactive structure, reducing the number of modules, making the network hierarchy clearer and the data flow more direct; thus, while ensuring prediction accuracy, the feasibility and engineering maintainability are improved, making the overall architecture simpler and more efficient.

[0030] In the aforementioned N-BEATSx plankton prediction method based on the fusion semantic model, in step S101, an observation dataset is established based on the collection and synchronization of multi-source environmental data and plankton data, thereby ensuring the temporal consistency and integrity of the data foundation for subsequent modeling. In step S102, multi-scale interpolation processing is performed on the observation dataset according to the interpolation strategy corresponding to different interval durations to obtain a target observation dataset containing target multi-source environmental data and target plankton data, thereby achieving data continuity and temporal scale alignment, providing a unified input for feature analysis. In step S103, cross-attention and residual operations are performed on the two types of data based on the improved N-BEATSx network of the fusion semantic model, so that the impact of environmental changes on the biological change rate sequence of plankton is expressed in a unified structure, and a prediction result of plankton change based on future time periods is generated. Based on this, the entire technical solution realizes the semantic fusion and dynamic prediction of environmental data and biological data, making the analysis of plankton change trends more accurate, coherent, and with stable interpretive logic.

[0031] In an exemplary embodiment, the observation dataset is subjected to multi-scale interpolation processing according to the interpolation strategies corresponding to different interval durations to obtain the target observation dataset, including steps S201 to S205.

[0032] Step S201: Divide the observation dataset into short-interval data segments, medium-interval data segments, and long-interval data segments according to different interval durations.

[0033] For example, firstly, in the observation dataset, the time index of each data sequence is read and the time difference between adjacent sampling points is calculated; then, based on the statistical distribution of the time difference, according to the pre-set boundary thresholds for short intervals, medium intervals and long intervals, data segments with higher sampling frequency and more uniformity are distinguished from data segments with sparse sampling and larger span, thereby dividing each data sequence in the observation dataset into short interval data segments, medium interval data segments or long interval data segments respectively.

[0034] The data segments are categorized into three types: short-interval data segments, medium-interval data segments, and long-interval data segments. Short-interval data segments correspond to high-frequency data under continuous sampling conditions; medium-interval data segments correspond to data with relatively stable sampling intervals but slight fluctuations; and long-interval data segments correspond to data with sampling interruptions or long time spans. For example, data sequences with sampling intervals less than 3 hours are considered short-interval data segments, data sequences with sampling intervals between 3 and 12 hours are considered medium-interval data segments, and data sequences with sampling intervals greater than 12 hours are considered long-interval data segments. Based on this, targeted interpolation strategies can be selected for data segments with different intervals, avoiding overfitting or trend shifts caused by using a uniform interpolation strategy.

[0035] Step S202: Based on the interpolation strategy corresponding to the short interval duration, the short interval data segment is interpolated according to the interval monotonicity constraint to obtain the target short interval data segment.

[0036] For example, interpolation based on interval monotonicity constraints is performed on short-interval data segments. This means that, due to the high sampling density of short-interval data, the changing trends between data points are relatively stable, and the main goal of interpolation is to maintain the original local change direction and continuity. Specifically, firstly, the local change direction of each interval in the short-interval data segment is calculated, i.e., the upward or downward trend between adjacent sampling points is determined, and this is used as the monotonicity constraint condition for the corresponding interval. Then, based on the monotonicity constraint conditions of each interval, an interpolation function is established within the data segment to ensure that the corresponding interpolation curve does not reverse direction between the original data points. In the specific calculation process, by determining the boundary derivative of each interval and the smooth connection of intermediate nodes, continuous completion of missing points is achieved. After interpolation is completed, the first-order derivative consistency check is performed on the corresponding interpolation curve to ensure that there are no abrupt changes or discontinuities at the junctions of adjacent intervals.

[0037] Based on this, interpolation is performed on each short-interval data segment to obtain the corresponding target short-interval data segment, which maintains the local trend characteristics of the original data numerically and achieves the continuity of the time series in terms of structure. It can be seen that the observed changes at short time scales usually reflect the slight fluctuations in the environment, and the error amplification can be effectively prevented by interpolation with interval monotonicity constraints.

[0038] Optionally, for short-interval data segments, the PCHIP (Piecewise Cubic Hermite Interpolating Polynomial) method can be used to maintain the monotonicity of data variation within the interval. Specifically, this method constructs an interval interpolation curve for each missing interval using the true values ​​of adjacent data points and their corresponding derivative information. By establishing a cubic polynomial relationship between adjacent data points, the interpolation result not only ensures continuous function values ​​at each endpoint but also guarantees the continuity of the first derivative, thus achieving a balance between smoothness and fidelity.

[0039] To avoid overshooting or non-physical oscillations at trend inflection points in the interpolation curve during the calculation process, the derivative values ​​of each node are first determined based on the direction of change of adjacent intervals: when the direction of change of the intervals on both sides of a node is inconsistent or there is a flat interval, the derivative of that node is set to zero to ensure a smooth transition of the curve at that point; when the direction of change of the intervals on both sides of a node is consistent, the node derivative is determined based on the trend of change of the preceding and following intervals to maintain the continuity of local changes. Subsequently, using these node derivatives and the corresponding real data points as constraints, a cubic polynomial function is constructed between each adjacent node to calculate any interpolation point within the missing interval, making the interpolation result numerically continuous, monotonic in form, and able to smoothly connect to the overall trend of the original data.

[0040] Step S203: Based on the interpolation strategy corresponding to the medium interval duration, interpolate the medium interval data segment based on the local slope weighting relationship to obtain the target medium interval data segment.

[0041] For example, interpolation based on local slope weighting is performed on medium-interval data segments. This means that, since the time span of medium-interval data segments is moderate, the overall trend is relatively flat but local differences are significant. Therefore, the main goal of interpolation is to balance the relationship between trend continuation and local correction. Specifically, firstly, the local change direction of each interval in the medium-interval data segment is calculated, i.e., the local slope of adjacent sampling points is calculated, and the rate of change of the local slope is used as a weighting coefficient to adjust the shape of the interpolation curve in the corresponding interval. For example, when the local slope difference is small, the shape of the interpolation curve tends to be linear and smooth; furthermore, when the local slope difference is large, a local weighting term is added at the corresponding position to appropriately adjust the shape of the interpolation curve to better reflect the data changes.

[0042] Therefore, through this weighted mechanism, the interpolation process not only considers the length of the time interval but also reflects the intensity of change between adjacent data points, thus avoiding the trend bias problem caused by simple linear interpolation. After interpolation is completed, the overall consistency of the resulting target medium-interval data segments is checked to ensure that the rate of change at the connection between each interval remains continuous.

[0043] Based on this, interpolation is performed on each medium-interval data segment to obtain the corresponding target medium-interval data segment, which numerically reflects the trend of change at the mesoscale and logically preserves the continuity of time and the coordination of local changes. It can be seen that by using local weighting, the interpolation curve is neither distorted nor overly smoothed, thus taking into account both the stability of mesoscale data and the fidelity of details.

[0044] Optionally, for data segments with medium intervals, the Akima (Akima Piecewise Cubic Hermite Interpolation) method can be used to effectively suppress overshoot while ensuring curve smoothness. This method is also based on a cubic polynomial function and is structurally similar to the PCHIP method, but there are significant differences in the derivative calculation constraints. Specifically, the Akima method determines the derivative value at each node by introducing a local slope weighting mechanism, that is, it comprehensively utilizes the rate of change information of the current node and multiple adjacent intervals before and after it, making the derivative calculation reflect more the local data distribution characteristics than the overall trend.

[0045] For example, when the rate of change differs significantly between adjacent intervals, the influence of abnormal slopes is automatically reduced, thus preventing unreasonable fluctuations in the interpolation curve at abrupt changes or inflection points. When the rate of change between adjacent intervals is similar, the nodal derivatives are smoothly transitioned through averaging to maintain curve continuity. Subsequently, these nodal derivative values ​​and the original nodal data are used to construct cubic polynomials within each interval, enabling numerical interpolation of missing data points within medium time intervals. This ensures that the interpolation curve is numerically continuous and naturally connected in shape, thereby more accurately reflecting the mesoscale temporal variation patterns.

[0046] Step S204: Based on the interpolation strategy corresponding to the long interval duration, the long interval data segment is interpolated according to the forward and backward weighted relationship to obtain the target long interval data segment.

[0047] For example, interpolation based on forward and backward weighting is performed on long-interval data segments. This is because long-interval data segments are characterized by large sampling intervals, numerous missing points, and uncertain trends. Therefore, the main goal of interpolation is to simultaneously utilize the data features from both the front and back sides for balanced calculations. Specifically, firstly, the time index of each sampling point in the long-interval data segment is extracted to identify time intervals with significant gaps, and the relative time position of each interpolation point within the corresponding interval is determined. Then, the time distance between each interpolation point and the sampling points at both ends of the corresponding interval is calculated. Specifically, the forward weight corresponding to an interpolation point is determined based on the time distance between that interpolation point and the previous sampling point within the corresponding interval to reflect the continuity of historical trends; the backward weight corresponding to an interpolation point is determined based on the time distance between that interpolation point and the next sampling point within the corresponding interval to reflect the guiding nature of future trends.

[0048] Next, for each interpolation point, the observations of the previous sampling point and the observations of the subsequent sampling point are weighted and superimposed according to their forward weights to obtain the interpolation value of that interpolation point within the corresponding interval. Then, the values ​​of each intermediate time period within the corresponding interval are smoothly distributed so that the interpolation result presents a gradual rather than abrupt characteristic. Thus, this calculation method avoids the trend drift problem caused by one-sided extrapolation, while ensuring that the interpolation result conforms to the data evolution logic on a long-term scale.

[0049] Based on this, interpolation is performed on each long-interval data segment to obtain the corresponding target long-interval data segment. Numerically, this achieves the completion of intervals with large time spans, and logically maintains the trend continuity and structural integrity of the overall data. It can be seen that by using bidirectional weighted constraints, reasonable filling of long-term missing data can be achieved, thereby maintaining the global continuity of the time series.

[0050] Optionally, for long-interval data segments, a forward-and-backward weighted PCHIP method can be used to fully utilize a wider range of temporal context information over a longer time scale. This method, while maintaining the monotonicity of traditional PCHIP interpolation, introduces a weighting mechanism in the time direction, allowing the interpolation process to simultaneously consider the influence of both forward and backward data points on the current interpolation point. Specifically, the forward weight reflects the continuity effect of historical data, while the backward weight reflects the moderating effect of future trends. The ratio of the two is dynamically determined based on the time interval length and the rate of data change, thus achieving a numerical balance of temporal information.

[0051] Therefore, through this weighting method, the interpolation function can naturally extend over a large time span, avoiding trend deviation or distortion caused by the dominance of one-sided data. In addition, the forward and backward weighting structure also simulates the inertial characteristics of time series to a certain extent, making the transition of the interpolation curve in long intervals smoother and the changes more coherent, thereby improving the stability and physical rationality of the interpolation results in long-term data prediction and reconstruction.

[0052] Step S205: Integrate the target short-interval data segment, the target medium-interval data segment, and the target long-interval data segment to obtain the target observation dataset.

[0053] For example, firstly, based on the time index, the short-interval data segments, medium-interval data segments, and long-interval data segments of each target are continuously arranged and / or superimposed on the time axis according to the actual observation order, thereby forming an overall time series structure, ensuring that the data segments are connected in time and without any omissions; then, a boundary smoothing operation is performed at the junctions of the data segments, and the numerical changes at the junctions are made continuous and consistent through local difference adjustments, avoiding abrupt changes or jumps between interpolation segments; next, the integrated data is subjected to an overall consistency check, checking whether the changing trends between the data segments conform to the statistical laws of the original data. If abnormal deviations are found, local weighted corrections are used to ensure that the overall trend remains consistent.

[0054] Therefore, the final target observation dataset achieves complete coverage in the temporal dimension and maintains multi-scale consistency in the numerical dimension. This dataset simultaneously includes multi-source environmental data and plankton data, providing a unified, continuous, and aligned input for subsequent semantic analysis and prediction. Based on this, it ensures that data from different time scales are structurally organically integrated, enabling a continuous transition from local smoothness to overall consistency in the entire interpolation process, resulting in a final data outcome with temporal integrity and logical coherence.

[0055] In this embodiment, in step S201, the observation dataset is divided according to the interval duration, thereby distinguishing the data structure at different time scales and providing a basis for targeted interpolation strategies. In step S202, short-interval data segments are interpolated based on interval monotonicity constraints to maintain consistent local change directions and avoid trend distortion of high-frequency data. In step S203, medium-interval data segments are interpolated based on local slope weighting relationships to balance data smoothness and fidelity of local details. In step S204, long-interval data segments are interpolated based on forward and backward weighting relationships to achieve smooth transitions and maintain overall trend continuity over large time spans. In step S205, the interpolated data segments are integrated to ensure that the integrated data is continuous and consistent on the time axis and has a coordinated trend. Based on this, the entire technical solution achieves unified completion and continuity maintenance of the observation dataset across multiple time scales, providing a stable, complete, and logically consistent time-series data foundation for subsequent analysis.

[0056] In an exemplary embodiment, before dividing the observation dataset into short-interval data segments, medium-interval data segments, and long-interval data segments according to different interval durations, the method further includes step S301: integrating the target short-interval data segments, target medium-interval data segments, and target long-interval data segments to obtain the target observation dataset, including steps S302 to S303.

[0057] Step S301: Based on the statistical distribution characteristics of the observation dataset, outlier data in the observation dataset is excluded to obtain an outlier-excluded observation dataset, which is then used to divide the dataset into short-interval data segments, medium-interval data segments, and long-interval data segments.

[0058] For example, anomaly removal processing is performed on the original observation dataset to ensure the stability and accuracy of subsequent interpolation and data segmentation. Specifically, firstly, based on the statistical distribution characteristics of the data, the numerical range, rate of change, and temporal continuity of each sampling point are analyzed; that is, the mean and standard deviation of each variable in the historical distribution are calculated, and a threshold range is set to determine whether there are any abnormal deviations. For values ​​that exceed the set threshold range or show abrupt changes in the time series, the degree of abnormality is determined through contextual comparison, and they are removed from the original observation dataset. In addition, to avoid accidental deletion, the data trend of adjacent time periods is also considered during the exclusion process to ensure the rationality of the anomaly identification.

[0059] Based on this, after anomaly removal, the retained data remains continuous in time and exhibits statistically consistent fluctuation characteristics in terms of value. It can be seen that outlier data often causes trend distortion or error propagation in subsequent interpolation calculations. Therefore, cleaning in advance can improve the stability of the data structure and lay a reliable numerical foundation for the subsequent division of short, medium and long interval data segments.

[0060] Step S302: According to the time index of each data segment, the target short interval data segment, the target medium interval data segment, and the target long interval data segment are time-aligned to obtain the fused dataset.

[0061] For example, firstly, time labels are extracted from each data segment to determine the start and end positions of each segment on the timeline. Since data segments at different time scales are processed independently during interpolation, their time boundaries need to be realigned to ensure overall temporal consistency. Based on this, the positions of each data segment on the timeline are adjusted according to its time span and sampling resolution, so that short-interval data segments are embedded into the detailed positions of medium-interval segments, and long-interval data segments cover the overall trend range. On this basis, all data segments are overlaid using time index mapping, creating a continuous, fused structure in the time dimension.

[0062] It is evident that the time steps and coverage vary across different interpolation scales. Without alignment, time misalignment or overlap can easily occur. Therefore, time synchronization ensures the overall structural coherence of the data. Based on this, the final fused dataset maintains complete coverage in the time dimension and achieves a unified representation of multi-scale data in the numerical dimension.

[0063] Step S303: According to the preset sliding window length, the fused dataset is subjected to smoothing filtering to obtain the target observation dataset.

[0064] For example, a smoothing filter based on a sliding window length is performed on the fused dataset to eliminate local fluctuations and discontinuities generated during interpolation. Specifically, firstly, a smoothing interval is determined according to a preset window length, so that the value at each time point is obtained by a weighted average of its neighboring time points, thereby achieving a smooth transition in the time dimension. During calculation, the sliding window moves across the entire fused dataset, gradually updating the value at each time point, so that the output sequence maintains the overall trend while reducing the impact of local noise. Furthermore, the window length is set based on the sampling frequency and observation period to ensure that the smoothing degree can filter out minor fluctuations without weakening the main trend features. In addition, after the smoothing filter, transition smoothing is performed at the intersection of different data segments to avoid data jumps.

[0065] As can be seen, although the interpolation results fill in the data gaps, there are still slight fluctuations at the intersection of data segments at different time scales. Smoothing with a sliding window can unify the data structure and improve the usability and consistency of the overall sequence, thereby forming a target observation dataset with stable time series characteristics. It is continuous in time, stable in numerical value, and retains the overall characteristics of the original observation in statistical structure.

[0066] In this embodiment, in step S301, anomaly removal processing is performed based on the statistical distribution characteristics of the observation dataset to remove distorted or abrupt values, ensuring the consistency and reliability of the data at the numerical level. In step S302, alignment is performed based on the time index of each data segment, thereby forming a continuously covered fused dataset of short, medium, and long interval data on the time axis, ensuring structural coordination between data at different time scales. In step S303, smoothing filtering is performed on the fused dataset according to the preset sliding window length, thereby reducing local fluctuations at the interpolation transition and maintaining the stability of the overall trend. Based on this, in the entire technical solution, through the coordinated processing of anomaly removal, time alignment, and smooth fusion, the observation data is completely reconstructed from its original distribution to a continuous time, making the multi-scale data structure more stable, unified, and possessing continuous temporal characteristics.

[0067] In an exemplary embodiment, in an improved N-BEATSx network that pre-integrates semantic models, cross-attention and residual operations are performed on target multi-source environmental data and target plankton data to obtain prediction results based on the change in plankton in the current environment, including steps S401 to S403.

[0068] Step S401: Sort the target multi-source environmental data according to semantic importance to obtain each semantic environmental factor.

[0069] Semantic importance represents the degree of correlation between various environmental factors and changes in plankton populations in historical observations. It is used to determine the reference priority of different environmental factors before entering the network, that is, to sort them according to the driving strength of environmental factors in the historical changes in the number of organisms for subsequent structural calls.

[0070] Semantic environment factors refer to a sequence of environment variables with priority attributes obtained based on semantic importance ranking. They are used to participate in subsequent feature operations in a fixed order in the network structure. For example, semantic environment factors with higher ranking are placed at the beginning of the structure to influence the generation of features on a larger scale.

[0071] For example, in the target multi-source environmental data, each environmental factor is sorted at the semantic level according to a pre-set semantic judgment rule. That is, the sorting process is not based on the numerical value, but on the semantic hierarchy based on the correlation between the environmental factor and the changes in plankton in historical observations, so that each environmental factor is given a clear semantic priority before entering the subsequent calculation structure.

[0072] In other words, changes in plankton are often driven by a variety of environmental factors. The importance of different environmental factors varies in different environmental scenarios. If all environmental factors are entered into subsequent structures with the same priority, the role of key variables will be diluted. However, by sorting by semantic importance, the reference priority of different environmental factors can be clearly defined at the network input, so that subsequent structures can call the corresponding environmental data in a sorted manner when performing interactive operations.

[0073] In this process, the synchronicity, responsiveness, and deviation of various environmental factors in the target multi-source environmental data with changes in phytoplankton abundance at different historical stages are analyzed to determine the reference value of the environmental factor in future prediction tasks. For example, temperature and phytoplankton abundance show a highly synchronized rise and fall relationship in the same season of multiple years, that is, the phytoplankton abundance increases when the temperature rises and decreases significantly when the temperature falls. In contrast, salinity does not change synchronously with phytoplankton abundance in most periods; even when salinity fluctuates greatly, the phytoplankton abundance remains stable. Therefore, by comparing this synchronicity and deviation, it can be determined that temperature has higher reference value in future predictions, while salinity is relatively less important.

[0074] Finally, after sorting, each environmental factor in the target multi-source environmental data is converted into a semantic environmental factor. This semantic environmental factor retains the time-series structure of the original environmental data and also carries semantic priority, giving it a clear position to participate in the subsequent network.

[0075] Step S402: Based on the preset double residual structure in the improved N-BEATSx network and the order of each semantic environment factor, determine the double residual block corresponding to each semantic environment factor in the double residual structure.

[0076] For example, in order for environmental factors to play an effective role in the double residual structure, it is necessary to determine the position of the corresponding double residual block in the double residual structure according to the ranking result of each semantic environmental factor. This allows the semantic environmental factors ranked earlier to obtain references closer to the input end in the structure, so that their influence covers more levels; the semantic environmental factors ranked later are placed in the residual blocks further back, so as to supplement the detailed influence at the local level of the network.

[0077] In other words, since the double residual structure itself forms a sequence expression in a progressive manner, the scope of different double residual blocks in the whole network is not the same. If the semantic ordering is not mapped to the structural level, the transmission path of environmental data in the network will lack order and it will be difficult to reflect the semantic differences of the environmental factors themselves.

[0078] This mapping method enables semantic environmental factors to obtain fixed and traceable positions in the double residual structure, allowing the network to always reference environmental data in a predetermined order during processing, thus forming a dual constraint of semantic and structural correspondence.

[0079] Step S403: Based on the cross-attention mechanism in each double residual block, perform cross-attention and residual operations on each semantic environmental factor and the target plankton data to obtain the prediction results based on the change in plankton under the current environment.

[0080] Among them, the cross-attention mechanism represents an operation method for establishing dynamic reference relationships between two types of sequences, which enables the biorate of change sequence of plankton to focus on the part of environmental data most relevant to its changes during the calculation process.

[0081] For example, starting from the initial layer of the dual residual structure, the most basic environmental impact reference relationships are established first, allowing semantic environmental factors to be dynamically correlated with the biological change rate sequence of phytoplankton in the target phytoplankton data within the same computational unit, thereby generating a preliminary expression of future change directions. Subsequently, when the structure moves to the next level, the input content no longer completely follows the initial sequence form, but instead supplements the output of the previous level with differentiated information, so that the input obtained at each level contains the change structure from the previous level, thus forming a progressive fusion path; that is, phytoplankton change often exhibits cumulative and phased characteristics, and each layer of the structure needs to understand the role of environmental factors at different depths of abstraction, therefore the input content must be gradually updated as the layers advance.

[0082] Within each level, cross-attention operations guide dynamic referencing between two types of inputs, enabling inputs related to phytoplankton abundance changes to focus on key segments of the corresponding semantic environmental factors at that level, extracting environmental impacts consistent with the current level's scope of interest. Next, residual operations maintain the continuity of the base sequence within the same structure, ensuring that newly generated expressions retain the overall trend of the original sequence after superposition. As computation unfolds sequentially across multiple levels, the structure gradually forms a composite expression that combines "environmental impact interpretation" and "historical structure preservation." This results in a sequence that reflects the phytoplankton change trend under current environmental conditions—a numerical sequence of future changes—providing directly usable outputs for subsequent ecological analysis.

[0083] In this embodiment, in step S401, semantic environmental factors are obtained by sorting the target multi-source environmental data, so that each environmental factor has a clear semantic importance order before entering the network, thereby ensuring that the subsequent structure has a clear priority when referencing environmental data; in step S402, according to the hierarchical mapping relationship of semantic environmental factors in the double residual structure, different environmental factors play an influence range that matches their semantic importance in structures at different depths, thereby forming an ordered hierarchical fusion path; in step S403, semantic environmental factors and target plankton data are fused layer by layer according to cross-attention and residual operation to generate prediction results, so that the prediction expression maintains the continuity of temporal structure while introducing environmental influences, and the plankton prediction process has structural consistency, interpretability, and prediction reliability.

[0084] In an exemplary embodiment, the method includes steps S501 and S601.

[0085] Step S501: If the current double residual block represents the first double residual block in the double residual structure, then the target plankton data and the corresponding semantic environmental factors are respectively used as the inputs of the current double residual block, and the predicted residual sequence of the current double residual block for future residuals and the regression sequence for the number of plankton are respectively used as the outputs of the current double residual block.

[0086] Step S601: If the current double residual block represents a double residual block other than the first double residual block in the double residual structure, then the residual sequence between the target plankton data and the regression sequence output by the previous double residual block, and the corresponding semantic environmental factors are respectively used as the inputs of the current double residual block, and the predicted residual sequence of the double residual block for future residuals and the regression sequence for the number of plankton are respectively used as the outputs of the current double residual block.

[0087] For example, on the one hand, if the current double residual block is in the first layer of the entire double residual structure, its input comes directly from the most basic original sequence, that is, the biodiversity change rate sequence in the target plankton data is input into this layer, enabling this layer to extract relevant trend and fluctuation features from the most fundamental historical changes in biodiversity. Simultaneously, the semantic environmental factors corresponding to this layer are introduced as another input, allowing environmental change to play a participating role from the initial stage of sequence modeling. Based on this input method, the first layer establishes the initial expression of the environment-biodiversity relationship, laying a unified baseline for subsequent layer-by-layer progression.

[0088] On the other hand, if the current double residual block is in a subsequent layer of the entire double residual structure, its input no longer directly uses the original biological rate of change sequence, but is based on the residual sequence formed between the regression output of the previous layer and the biological rate of change sequence. This allows the layer to identify insufficiently expressed bias information based on the results of the previous layer. Simultaneously, the semantic environmental factors corresponding to this layer are introduced as another input, ensuring that the layer maintains consistency in referencing environmental changes when compensating for biases. Based on this input method, subsequent layers can gradually improve the error information left by the preceding layers, forming a continuous correction path in the depth dimension of the entire structure.

[0089] Furthermore, after any dual residual block completes its internal operations, its output consists of two parts: first, a predictive residual sequence for future periods, representing the current layer's biased estimate of future trends, which is continuously superimposed in the vertical hierarchy to ultimately form a complete predictive residual structure; second, a regression sequence for plankton abundance, expressing the regression output of the current layer's biomass after integrating its input information, thus providing a basis for the differential input of subsequent layers. Therefore, through this dual-output mechanism, the entire residual hierarchy can form a consistent expression of future changes in biomass abundance based on continuous correction.

[0090] In this embodiment, in step S501, the target plankton data and corresponding semantic environmental factors are directly used as the first-layer input, enabling the network to establish the most basic environment-biological relationship in the initial stage, thereby ensuring that subsequent layers have a unified and reliable modeling starting point. In step S601, the residual sequence formed by the regression output of the previous layer and the target plankton data, combined with the semantic environmental factors, is used as input, enabling subsequent layers to progressively compensate for the biases not fully expressed in the previous layer, thereby improving the overall trend structure layer by layer and maintaining the consistency of environmental impact. Based on this, in the entire technical solution, the plankton prediction process forms a continuous, stable, and interpretable residual progressive structure, thereby improving the accuracy and structural consistency of change prediction.

[0091] In an exemplary embodiment, the step of performing cross-attention and residual operations on each semantic environmental factor and the target plankton data according to the cross-attention mechanism in each double residual block to obtain the prediction result based on the change in plankton in the current environment includes steps S701 to S703.

[0092] Step S701: The semantic environmental factors received by the current double residual block are used as the key vector, and the other input received by the double residual block is used as the query vector. The key vector and the query vector are cross-attentioned in a preset cross-attention layer to obtain a value vector based on the semantic environmental factors. If the current double residual block represents the first double residual block in the double residual structure, the other input received is the target plankton data. If the current double residual block represents a double residual block other than the first double residual block in the double residual structure, the other input received is the residual sequence.

[0093] Here, the key vector represents a sequence of features formed by semantic environmental factors, which is used to provide the location basis of the queried information in cross-attention operations, such as as the source of the time segment of interest when modeling the correlation between environmental changes and changes in plankton abundance.

[0094] The query vector represents a feature representation sequence formed by the biological change rate sequence or residual sequence of plankton. It is used to actively search for environmental information related to itself in the cross-attention operation, such as selecting the corresponding environmental feature content from the key vector when identifying the cause of change in a certain period of time.

[0095] The value vector represents the fused feature obtained after cross-attention calculation of the key vector and the query vector. It is used to express the effect of environmental factors on the current plankton population and serves as the input basis for trend bias and population prediction in subsequent residual calculations.

[0096] The cross-attention layer represents a feature interaction structure used to establish dynamic reference relationships between two types of sequences. Specifically, it is used to select the information most relevant to the current sequence position from the key vector based on the query vector and generate a value vector for subsequent residual operations.

[0097] For example, firstly, based on the hierarchical position of the current double residual block in the overall structure, the two types of input formats that the layer needs to receive are determined. Then, the semantic environmental factor is used as the key vector of the layer, and the other input is used as the query vector, so that both are sent to a preset cross-attention layer. In this process, if the current double residual block is located in the first layer of the structure, the query vector comes from the biological change rate sequence in the target plankton data, enabling the first layer to establish the initial information connection between environmental factors and biological change based on the original sequence; if the current double residual block is located in a subsequent layer, the query vector comes from the residual sequence formed between the regression sequence and the biological change rate sequence of the previous layer, enabling the layer to refocus on the location of the environmental factors that can better explain the source of the deviation around the deviation structure of the previous layer.

[0098] Furthermore, the cross-attention operation is performed by calculating the correspondence between the key vector and the query vector within the cross-attention layer. This allows the query vector to establish a weighted reference path across different time segments of environmental factors, thereby generating a value vector that reflects the characteristics of environmental impact. Essentially, since phytoplankton changes are driven by environmental factors, and the degree of their impact varies across different time segments, it is necessary to extract the content that best strengthens the explanatory power of the current layer from the environmental data through interactive referencing. Thus, the final value vector serves as the output of the cross-attention stage, constituting the core feature representation of this layer before entering subsequent residual operations.

[0099] For example, if the number of plankton decreases significantly over a certain period of time, while the temperature rises suddenly at the same location, the cross-attention layer will cause the query vector to pay more attention to the corresponding temperature change segment, thereby forming a feature expression in the value vector that the temperature change affects the downward trend.

[0100] Step S702: Based on the preset feedforward neural network and linear layer, the value vector is transformed to obtain the predicted residual sequence of the current double residual block for future residuals and the regression sequence for the number of plankton, so as to realize the residual operation of the double residual block.

[0101] Among them, the feedforward neural network represents a mapping structure used to perform nonlinear transformation on fused features, and is used to reorganize the internal expression of the value vector so that it can be used to generate trend bias and quantity prediction.

[0102] The linear layer represents a mapping structure that projects features onto the target output dimension, used to generate predictive residual sequences or regression sequences.

[0103] For example, the value vector generated in the aforementioned steps is fed into a preset transformation structure to complete the residual operation of the layer; wherein, the transformation structure consists of a feedforward neural network for feature transformation and two linear layers for generating different outputs, and its overall function is to reorganize the value vector obtained in the cross-attention stage so that it can respectively represent the estimation of future residuals and the regression expression of plankton quantity.

[0104] Specifically, the value vector is first processed in the transformation structure into an internal representation suitable for expressing the trend of numerical change, so that the temporal hierarchy of the data can be rearranged so that the subsequent output can maintain a change structure consistent with the time dimension. Then, the internal representation is sent to two output paths, one for generating a predicted residual sequence for future residuals and the other for generating a regression sequence for the number of plankton.

[0105] Step S703: Combine the predicted residual sequences output by each double residual block to perform difference calculation on the target plankton data to obtain the prediction results based on the change in plankton in the current environment.

[0106] For example, firstly, the predicted residual sequences of each layer are aligned in the time dimension so that the predicted bias calculated by each layer can be superimposed on the original biological rate of change sequence at the same time position. Subsequently, the system accumulates the predicted residual sequences of each layer in structural order, so that the information of step-by-step correction in the multi-layer structure can form a complete trend adjustment path. The reason for adopting a step-by-step superposition method is that the design purpose of the multi-layer double residual structure is to improve the overall prediction by continuously supplementing the bias information, rather than relying on the direct estimation of a single layer. Therefore, the final prediction result must include the cumulative effect of the residuals of each layer.

[0107] Finally, after the residual superposition is completed, the cumulative deviation is interpolated with the original biological change rate sequence to obtain the final prediction sequence reflecting the future change trend. This prediction sequence expresses the magnitude and direction of future plankton change in a continuous numerical form, serving as a complete output of the plankton change pattern under current environmental conditions and providing a directly usable time series expression for subsequent analysis and environmental assessment.

[0108] Based on this, the network retains the hierarchical progressive sequence decomposition method of the original N-BEATSx in its overall structure, allowing trends, fluctuations, and seasonal structures to be decomposed layer by layer while maintaining interpretability, thus providing a clear source of composition for the prediction results. Simultaneously, semantic cross-attention is incorporated into each layer, enabling the dynamic establishment of the interaction between semantic environmental factors and the biorate of change in plankton during the computation process. This allows the network to identify the specific impacts of environmental changes on different time segments. Therefore, the contribution of environmental factors is no longer implicit in the black-box structure but influences the prediction results in a traceable and interpretable manner, giving the overall network both structured decomposition capabilities and a transparent expression of environmental relationships.

[0109] Furthermore, the innovative approach of using key vectors, query vectors, and value vectors for cross-attention operations in this network scenario stems from the fact that phytoplankton changes are driven by both their own historical structure and the asynchronous influence of multiple environmental factors, with no fixed temporal correspondence between the two. Therefore, conventional feature concatenation or linear combination cannot establish a link between the environment and historical quantitative changes; instead, a structure capable of dynamically locating relevance across time is needed. By allowing the query vector to actively seek the most relevant environmental fragments from the key vectors and forming a fusion expression in the form of a value vector that can directly participate in residual calculation, the influence path of environmental factors can be continuously reinforced across multiple layers. Thus, this method of actively retrieving relevant environmental information based on semantic relationships and embedding it into a progressive residual structure is not naturally derived from traditional time series models. Its core lies in establishing a dynamic association mechanism across sequences, time periods, and semantics, thereby demonstrating its innovativeness.

[0110] In this embodiment, in step S701, a dynamic reference relationship between sequences is established based on the cross-attention operation between semantic environmental factors and another input, so that the impact of environmental changes on phytoplankton abundance changes is explicitly expressed in the form of associated vectors; in step S702, the characteristic vectors obtained in the previous steps are transformed and prediction residual sequences and regression sequences are generated respectively, so that the current layer can simultaneously provide the information basis for quantity estimation and trend deviation correction; in step S703, the difference between the prediction residual sequences of each layer is superimposed to form a continuous deviation information link, so that the final prediction result can reflect the comprehensive change after multi-layer correction; based on this, the unity of environmental association expression and trend change prediction is achieved in the entire technical solution, making the prediction of phytoplankton change both interpretable and continuous.

[0111] In one exemplary embodiment, Figure 2 A schematic diagram of the structure of a dual residual block of an improved N-BEATSx network that integrates semantic models is shown. Specifically, firstly, the biological change rate sequence is calculated separately from the regression sequence output by the previous dual residual block STACK(i-1) to obtain the residual sequence. Then, the residual sequence and the i-th semantic environmental factor in the semantic importance ranking are used as the input of the i-th dual residual block STACKi.

[0112] Next, in the cross-attention layer of the dual residual block STACKi, the residual sequence is used as the query vector and the semantic environmental factors are used as the key vector. The correspondence between the query vector and the key vector is calculated so that the query vector can establish a weighted reference path between different time segments of the key vector, thereby generating a value vector that reflects the characteristics of environmental impact.

[0113] Furthermore, the value vector is input into the feedforward neural network and processed into an internal representation suitable for expressing the trend of numerical change, so that the temporal hierarchy of the data can be rearranged so that the subsequent output can maintain a change structure consistent with the time dimension. Subsequently, the internal representation is fed into the prediction linear layer and the regression linear layer respectively. On the one hand, a linear mapping is performed based on the prediction linear layer to generate the prediction residual sequence of the current layer. On the other hand, independent regression calculation is performed based on the regression linear layer to generate the regression sequence of the current layer.

[0114] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0115] Based on the same inventive concept, this application also provides an N-BEATSx planktonic prediction system for implementing the N-BEATSx planktonic prediction method with fused semantic models as described above. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more N-BEATSx planktonic prediction system embodiments with fused semantic models provided below can be found in the limitations of the N-BEATSx planktonic prediction method with fused semantic models described above, and will not be repeated here.

[0116] In one exemplary embodiment, such as Figure 3 As shown, an N-BEATSx planktonic prediction system integrating semantic models is provided, including: an acquisition module 101, an interpolation module 102, and a prediction module 103, wherein: The acquisition module 101 is used to acquire the observation dataset under the current environment. The observation dataset includes multi-source environmental data and plankton data. Interpolation module 102 is used to perform multi-scale interpolation processing on the observation dataset according to the interpolation strategies corresponding to different interval durations, so as to obtain the target observation dataset, which includes target multi-source environmental data and target plankton data; The prediction module 103 is used to perform cross-attention and residual operations on target multi-source environmental data and target plankton data in an improved N-BEATSx network that pre-integrates semantic models, so as to obtain prediction results based on the change in plankton in the current environment.

[0117] Each module in the aforementioned N-BEATSx planktonic prediction system, which incorporates semantic models, can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, allowing the processor to invoke and execute the corresponding operations of each module.

[0118] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above embodiments.

[0119] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above embodiments.

[0120] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0121] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An N-BEATSx plankton prediction method with fused semantic models, characterized in that, The method comprises: obtaining an observation data set in a current environment, the observation data set comprising multi-source environmental data and plankton data; performing multi-scale interpolation processing on the observation data set according to interpolation strategies corresponding to different interval lengths, to obtain a target observation data set, the target observation data set comprising target multi-source environmental data and target plankton data; performing cross-attention operation and residual operation on the target multi-source environmental data and the target plankton data in an improved N-BEATSx network pre-fused with a semantic model, to obtain a prediction result of a plankton change amount in the current environment.

2. The method of claim 1, wherein, The method comprises: dividing the observation data set into short-interval data segments, medium-interval data segments and long-interval data segments according to different interval lengths; performing interpolation processing on the short-interval data segments based on interval monotonicity constraints according to an interpolation strategy corresponding to a short-interval length, to obtain target short-interval data segments; performing interpolation processing on the medium-interval data segments based on local slope weighting relationships according to an interpolation strategy corresponding to a medium-interval length, to obtain target medium-interval data segments; performing interpolation processing on the long-interval data segments based on forward and backward weighting relationships according to an interpolation strategy corresponding to a long-interval length, to obtain target long-interval data segments; integrating the target short-interval data segments, the target medium-interval data segments and the target long-interval data segments to obtain a target observation data set.

3. The method of claim 2, wherein, Before the observation data set is divided into short-interval data segments, medium-interval data segments and long-interval data segments according to different interval lengths, the method comprises: excluding abnormal data in the observation data set according to statistical distribution characteristics of the observation data set, to obtain an observation data set after abnormal data exclusion, for division into short-interval data segments, medium-interval data segments and long-interval data segments; The integration of the target short-interval data segments, the target medium-interval data segments and the target long-interval data segments to obtain a target observation data set comprises: aligning the target short-interval data segments, the target medium-interval data segments and the target long-interval data segments in time according to time indexes of the data segments, to obtain a fused data set; performing smoothing filtering processing on the fused data set according to a preset sliding window length, to obtain a target observation data set.

4. The method of claim 1, wherein, The cross-attention operation and residual operation on the target multi-source environmental data and the target plankton data in the improved N-BEATSx network pre-fused with the semantic model to obtain the prediction result of the plankton change amount in the current environment comprises: sorting the target multi-source environmental data according to semantic importance, to obtain semanticized environmental factors; determining double residual blocks corresponding to the semanticized environmental factors in a double residual structure in the improved N-BEATSx network, in combination with the sorting of the semanticized environmental factors. According to the cross-attention mechanism in each double residual block, cross-attention operation and residual operation are performed on each semantic environmental factor and the target plankton data to obtain the prediction result based on the plankton change amount in the current environment.

5. The method of claim 4, wherein, The method further comprises: If the current double residual block represents the first double residual block in the double residual structure, the target plankton data and the corresponding semantic environmental factor are taken as inputs of the current double residual block respectively, and the prediction residual sequence of the future residual and the regression sequence of the plankton quantity output by the current double residual block are taken as outputs of the current double residual block respectively.

6. The method of claim 4, wherein, The method further comprises: If the current double residual block represents a double residual block other than the first double residual block in the double residual structure, a residual sequence between the target plankton data and the regression sequence output by the last double residual block, and the corresponding semantic environmental factor are taken as inputs of the current double residual block respectively, and the prediction residual sequence of the future residual and the regression sequence of the plankton quantity output by the double residual block are taken as outputs of the current double residual block respectively.

7. The method according to claim 5 or 6, characterized in that, The method further comprises: The semantic environmental factor received by the current double residual block is taken as a key vector, and another input received by the double residual block is taken as a query vector, cross-attention operation is performed on the key vector and the query vector in a preset cross-attention layer to obtain a value vector based on the semantic environmental factor; wherein, if the current double residual block represents the first double residual block in the double residual structure, the other input received is the target plankton data, and if the current double residual block represents a double residual block other than the first double residual block in the double residual structure, the other input received is a residual sequence; According to the preset feedforward neural network and linear layer, the value vector is transformed to obtain the prediction residual sequence of the future residual and the regression sequence of the plankton quantity output by the current double residual block, so as to realize the residual operation of the double residual block; According to the prediction residual sequence output by each double residual block respectively, difference calculation is performed on the target plankton data to obtain the prediction result based on the plankton change amount in the current environment.

8. An N-BEATSx plankton prediction system with fused semantic models, characterized in that, The system comprises: An acquisition module is configured to acquire an observation data set in a current environment, wherein the observation data set comprises multi-source environmental data and plankton data; An interpolation module is configured to perform multi-scale interpolation processing on the observation data set according to an interpolation strategy corresponding to different interval lengths respectively to obtain a target observation data set, wherein the target observation data set comprises target multi-source environmental data and target plankton data; A prediction module is configured to perform cross-attention operation and residual operation on the target multi-source environmental data and the target plankton data in an improved N-BEATSx network pre-fused with a semantic model to obtain a prediction result based on a plankton change amount in the current environment. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images