N-BEATSx network prediction method fusing multi-source environmental factors

Abstract

The invention relates to an N-BEATSx network prediction method fusing multi-source environmental factors. The method comprises the following steps: acquiring multi-source environmental data and plankton observation data in a current environment; converting the multi-source environmental data into different environmental factor covariables, and converting the plankton observation data into a biomass sequence; and inputting each environment factor covariable and the biomass sequence into a preset improved N-BEATSx network, and calculating a contribution result of each environment factor covariable to the biomass sequence according to the improved N-BEATSx network to obtain a prediction result based on the plankton variation in the current environment. By adopting the method, systematicness and accuracy of plankton quantity change prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecological monitoring, in particular to a N-BEATSx network prediction method fusing multi-source environmental factors. BACKGROUND

[0002] In the technical field of ecological monitoring, the number of plankton is predicted according to environmental factors, so as to realize quantitative evaluation of dynamic changes of an ecological system. In a related prediction method, a single environmental variable or a fixed weight model is used to fit the number of plankton, however, the method ignores the time sequence coupling relationship between multi-dimensional environmental factors, so that the prediction accuracy is insufficient, the environmental response characteristics are not complete, and it is difficult to reflect the dynamic correlation of the real ecological process. SUMMARY

[0003] Therefore, it is necessary to provide a N-BEATSx network prediction method fusing multi-source environmental factors, a device, a computer equipment and a computer readable storage medium aiming at the above technical problems.

[0004] In a first aspect, the present application provides a N-BEATSx network prediction method fusing multi-source environmental factors, comprising: obtaining multi-source environmental data and plankton observation data under a current environment; converting the multi-source environmental data into different environmental factor covariates, and converting the plankton observation data into a plankton number sequence; inputting each environmental factor covariate and the plankton number sequence into a preset improved N-BEATSx network, calculating a contribution result of each environmental factor covariate to the plankton number sequence according to the improved N-BEATSx network, so as to obtain a prediction result based on a plankton change amount under the current environment.

[0005] In a second aspect, the present application further provides a N-BEATSx network prediction device fusing multi-source environmental factors, comprising: an obtaining module, configured to obtain multi-source environmental data and plankton observation data under a current environment; a conversion module, configured to convert the multi-source environmental data into different environmental factor covariates, and convert the plankton observation data into a plankton number sequence; an analysis module, configured to input each environmental factor covariate and the plankton number sequence into a preset improved N-BEATSx network, calculate a contribution result of each environmental factor covariate to the plankton number sequence according to the improved N-BEATSx network, so as to obtain a prediction result based on a plankton change amount under the current environment.

[0006] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above steps when executing the computer program.

[0007] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above steps.

[0008] The fusion multi-source environmental factor N-BEATSx network prediction method, device, computer device and computer readable storage medium described above, firstly, the multi-source environmental data is converted into different environmental factor covariates, and the plankton observation data is converted into biological quantity sequences, so as to realize the structured expression of the observation data and the unified modeling of the input features; secondly, the improved N-BEATSx network is used to jointly analyze and calculate the contribution results of the environmental factor covariates and the biological quantity sequences, so as to analyze the time dimension of the action process of the environmental factors, and jointly deduce the change trend of the biological quantity, so as to generate a prediction result with time sequence consistency and causal correlation; based on this, in the whole technical solution, the improved N-BEATSx network can be used to realize the dynamic correlation modeling between the multi-source environmental factors and the biological quantity change, so as to reflect the dynamic correlation of the real ecological process, thereby improving the systematicness and accuracy of the plankton quantity change prediction. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments or the related art, the drawings needed to be used in the embodiment or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 A flowchart of the N-BEATSx network prediction method fusing multi-source environmental factors in an embodiment; Figure 2 A flowchart of the N-BEATSx network prediction method fusing multi-source environmental factors in another embodiment; Figure 3 A structure diagram of the double residual block of the improved N-BEATSx network in an embodiment; Figure 4 A structure block diagram of the N-BEATSx network prediction device fusing multi-source environmental factors in an embodiment. DETAILED DESCRIPTION

[0011] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0012] In one embodiment, as shown in Figure 1 A method for predicting the number of plankton in a multi-source environment is provided. The method can be applied to a server, a terminal, or a system including a terminal and a server, and can be implemented through the interaction between the terminal and the server.

[0013] In step S101, multi-source environmental data and plankton observation data in the current environment are obtained.

[0014] The current environment refers to a specific water environment for which the number of plankton is to be predicted, such as a water area of a lake, river or ocean in a specific season.

[0015] The multi-source environmental data refers to different types of environmental observation data obtained in the current environment, such as water temperature, salinity, dissolved oxygen content, pH, turbidity and chlorophyll content, which represent various environmental factors in multi-dimensional physical and chemical descriptions. The plankton observation data refers to the number of different types of plankton in the current environment, such as copepods, echinoderms, nostoc, jellyfish, noctiluca, shrimps, brown algae, arrow worms and tail sea slugs.

[0016] In step S102, the multi-source environmental data is converted into different environmental factor covariates, and the plankton observation data is converted into a biological quantity sequence.

[0017] The environmental factor covariates refer to a set of feature variables formed by standardizing, time slicing and encoding conversion of the multi-source environmental data, which are used to quantify the input influence of different environmental factors on the change of biological quantity, such as the feature components of temperature change rate, salinity gradient or light cycle parameters after processing.

[0018] The biological quantity sequence refers to a continuous numerical sequence formed by reconstructing the plankton observation data on the time axis, which is used to describe the change of the number of plankton in continuous time, such as the total cell number sequence of phytoplankton counted by day or the individual density sequence of zooplankton counted by hour.

[0019] Exemplarily, in one aspect, the multi-source environmental data is processed to obtain different environmental factor covariates. Specifically, after time synchronization of the multi-source environmental data, the value characteristics and dimension attributes of various environmental factors are identified, and normalization and unit standardization operations are performed thereon to eliminate inconsistencies caused by differences in sampling frequency, unit scale and measurement accuracy. Subsequently, the standardized data is divided into consecutive time segments in chronological order, each time segment representing the environmental state within a specific time window, and the values of various environmental factors within each time window are serialized, combined and encoded to form corresponding feature vectors of various environmental factors in a unified calculation dimension. Further, to ensure that the correlation between environmental factors can be kept analyzable in subsequent calculations, normalization, offset correction and scale balancing are further performed on the feature vectors corresponding to various environmental factors, thereby generating an environmental factor covariate set with time consistency and calculability; each environmental factor covariate in the environmental factor covariate set corresponds to a specific type of environmental factor to describe the input characteristics of the corresponding environmental change on the ecological response.

[0020] On the other hand, the plankton observation data is processed to obtain a biological quantity sequence. Specifically, first, the integrity of the plankton observation data is detected and outliers are removed to ensure the accuracy and consistency of the biological quantity information and time labels in the corresponding relationship. Subsequently, the biological quantity is sorted and resampled with time index as the main axis, so that the data presents a continuous and smooth change structure in the time dimension. In addition, for the time discontinuity introduced by different sampling periods or sampling errors, numerical completion is further performed by interpolation to ensure that each time period has effective biological quantity data support. After the above processing, the observation data of various plankton is reconstructed on the time axis to form a biological quantity sequence with time as the independent variable and plankton quantity as the dependent variable; the biological quantity sequence not only reflects the quantity change characteristics of a single type of plankton in the time dimension, but also integrates the overall quantity change characteristics of various plankton in different time scales, providing complete and continuous biological quantity data support for subsequent analysis.

[0021] Based on this, through the above two aspects of processing, the transition from the original observation data to the calculation input format is logically completed, so that the environmental factor covariates and the biological quantity sequence present calculability in structure, laying a foundation for subsequent joint calculation.

[0022] Step S103, input each environmental factor covariate and the biological quantity sequence into the preset improved N-BEATSx network, calculate the contribution result of each environmental factor covariate to the biological quantity sequence according to the improved N-BEATSx network, and obtain the prediction result based on the plankton change amount under the current environment.

[0023] Among them, the improved N-BEATSx network represents a multi-layer residual prediction structure based on time series decomposition, which is used to jointly analyze the time sequence mapping relationship between the environmental factor covariate and the biological quantity sequence, so as to output the prediction result of the future biological quantity change.

[0024] Among them, the contribution result of each environmental factor covariate to the biological quantity sequence represents the relative influence degree of different environmental factors in the biological quantity change process, which is used to quantitatively describe the action strength and direction of each environmental factor on the increase and decrease of plankton quantity.

[0025] Exemplarily, based on the residual calculation structure of the improved N-BEATSx network, the environmental factor covariate and the biological quantity sequence obtained in the foregoing step are jointly calculated to determine the contribution result of each environmental factor covariate to the biological quantity sequence and generate the prediction result. Specifically, first, after the environmental factor covariate and the biological quantity sequence are indexed and aligned in the time dimension, they are input into the network for residual progressive calculation; wherein, in the residual progressive calculation process, the environmental factor covariate and the biological quantity sequence are jointly used as the input feature set, and the feature mapping and residual update are performed under the unified time index, so as to keep the synchronization transmission of environmental change information and biological quantity change information in the network structure.

[0026] Further, the network is composed of multiple layers of residual blocks in cascade, each residual block is stacked by linear combination and nonlinear transformation, which decomposes the input environmental factor covariate layer by layer to calculate the feature response of the environmental factor on different time scales, and further obtains the residual information transmitted to the next layer according to the feature response, so as to realize the recursive extraction of time features in the deep structure. Further, with the progress of residual progressive calculation, the network parameters of each layer are iteratively updated by the preset back propagation algorithm, so that the network gradually approaches the optimal mapping relationship between the environmental factor covariate and the biological quantity sequence.

[0027] Finally, the output layer of the network calculates the contribution result of each environmental factor covariate to the biological quantity sequence based on the feature response accumulated by the multiple layers of residual blocks, and generates the biological quantity prediction result in the future time window according to the contribution result; the prediction result is in the form of time series, so that the prediction value of each time point corresponds to the contribution result of each environmental factor covariate respectively, so as to realize the structured and quantitative expression of the environmental change to the biological quantity change.

[0028] Based on this, through the hierarchical decomposition mechanism of the improved N-BEATSx network, the action process of the environmental factor is analyzed in the time dimension, and the change trend of the biological quantity is jointly deduced, so as to generate a prediction result with time consistency and causal correlation, which can be directly used as a basis for subsequent verification, comparative analysis and ecological trend evaluation.

[0029] In the above N-BEATSx network prediction method of fusing multi-source environmental factors, first, the multi-source environmental data is converted into different environmental factor covariates, and the plankton observation data is converted into biological quantity sequences, so as to realize the structured expression of the observation data and the unified modeling of the input features. Further, the improved N-BEATSx network is used to jointly analyze the environmental factor covariates and the biological quantity sequences and calculate the contribution results, so as to analyze the action process of the environmental factor in the time dimension, and jointly deduce the change trend of the biological quantity, thereby generating a prediction result with time consistency and causal correlation. Based on this, in the entire technical solution, the improved N-BEATSx network can be used to realize dynamic correlation modeling between multi-source environmental factors and biological quantity changes, so as to reflect the dynamic correlation of the real ecological process, thereby improving the systematicness and accuracy of the plankton quantity change prediction.

[0030] In one exemplary embodiment, the contribution results of the environmental factor covariates to the biological quantity sequences are calculated based on the improved N-BEATSx network to obtain a prediction result based on the plankton change amount under the current environment, including steps S201 to S203.

[0031] In step S201, the long-term correlation between each environmental factor covariate and the biological quantity sequence is calculated to obtain the sorting results of each environmental factor covariate according to the corresponding long-term correlation.

[0032] Exemplarily, the long-term correlation between each environmental factor covariate and the biological quantity sequence is calculated to construct the influence degree of each environmental factor in the time dimension. Specifically, first, the time sequence corresponding to each environmental factor is extracted from each environmental factor covariate, ensuring that its time index is strictly aligned with the biological quantity sequence, thereby maintaining the time consistency between the data. Subsequently, the long-term correlation of each environmental factor covariate and the biological quantity sequence is calculated, and the statistical dependence relationship thereof in a long time scale is calculated to capture the continuous influence characteristics of the environmental factor on the biological quantity change. In this process, the long-term correlation not only reflects the direct action strength of the environmental factor, but also embodies the delay and persistence of the cumulative effect of the environmental factor in the time sequence.

[0033] Next, in order to ensure the comparability between different environmental factors, the correlation results are normalized to express the correlation values in a unified numerical interval. Finally, the normalized correlation results are sorted, and the environmental factors with higher long-term correlation are arranged in the front to reflect their dominant influence on the change of biological quantity. Based on this, a set of sorting results corresponding to the environmental factor covariates arranged from high to low according to the long-term correlation is obtained to reflect the importance of each environmental factor in the overall processing of the network.

[0034] In step S202, according to the double residual structure preset in the improved N-BEATSx network, the sorting results are combined to determine the corresponding double residual blocks of each environmental factor covariate in the double residual structure.

[0035] For example, the sorting results corresponding to the environmental factor covariates obtained according to the foregoing steps are input, and the mapping positions of the environmental factor covariates in the network are determined in combination with the double residual structure preset in the improved N-BEATSx network. Specifically, first, according to the arrangement order in the sorting results, the environmental factor covariates with higher long-term correlation are assigned to the double residual blocks with shallower levels in the residual structure to preferentially extract the main trend features; and the environmental factor covariates with lower long-term correlation are assigned to the double residual blocks with deeper levels in the residual structure to capture subtle or secondary influence features in subsequent progressive calculations, thereby realizing hierarchical analysis and step-by-step fusion.

[0036] In this process, through the preset structure index mapping mechanism, the sorting index of each environmental factor covariate is matched with the number of the double residual block, and a one-to-one correspondence is established; then, the input channels of each double residual block are initialized in the double residual structure to receive the corresponding environmental factor covariate input signal in subsequent calculations; in addition, in order to ensure the continuity of residual calculation, a cross-layer connection path is set inside the double residual structure, so that different double residual blocks can share part of the feature response, thereby realizing multi-layer progressive time decomposition.

[0037] Based on this, after the mapping and matching process is completed, the correspondence between the environmental factor covariates and the double residual blocks is formed, which not only determines the network flow direction of the input features, but also determines the level and range of action of each environmental factor in the network participating in the residual calculation; based on this, through this structured mapping, the hierarchical configuration of the environmental factor covariates in the double residual structure is completed.

[0038] In step S203, the contribution of each double residual block to the biological quantity sequence is calculated respectively to obtain the prediction result of the change amount of plankton under the current environment.

[0039] Exemplarily, based on the correspondence between the environmental factor covariates and the double residual blocks, residual progressive calculation is performed on each double residual block to determine the contribution results of different environmental factor covariates to the biological quantity sequence. Specifically, first, each double residual block simultaneously receives its corresponding environmental factor covariate and the time-aligned biological quantity sequence as input signals. During the calculation process, the input signals are linearly combined and differenced to extract the explainable part and the unexplained part of the environmental factor to the biological quantity change in the calculated feature response. The explainable part represents the response relationship of the environmental factor to the biological quantity sequence at a specific time scale, and is used to measure its direct contribution to the overall quantity change trend. The unexplained part is defined as residual information, which represents the unexplained biological quantity change of the environmental factor.

[0040] Then, the residual information is passed to the next layer of double residual blocks, so that the subsequent layers can make supplementary calculations based on the unexplained part of the previous layer, realizing hierarchical decoupling and progressive modeling of different environmental factors. With the continuous calculation of multiple layers of double residual blocks, the feature responses output by each double residual block are obtained and quantified, which are used as the contribution results of each environmental factor covariate to the biological quantity sequence. The contribution results reflect the influence proportion of each environmental factor in the overall prediction, and reflect the comprehensive effect of the environmental factor on the biological quantity change.

[0041] Finally, in the output layer of the network, the contribution results of each environmental factor covariate to the biological quantity sequence are fused to generate the prediction value of the biological quantity change at each time point in the future period, and then the prediction value at each time point is integrated into the prediction result of the plankton quantity change in the future period under the current environment. Based on this, the network completes progressive residual calculation and prediction output under the joint driving of multiple environmental factors, and realizes the structured mapping between the environmental factor effect and the biological quantity change.

[0042] In this embodiment, first, according to the calculation and sorting of the long-term correlation between each environmental factor covariate and the biological quantity sequence, the dominant influence degree of different environmental factors is distinguished in the time dimension, and a quantitative basis is provided for the subsequent hierarchical distribution of residual structure; secondly, according to the double residual structure in the improved N-BEATSx network, the environmental factor covariate is combined with the corresponding double residual block to establish a corresponding relationship, so as to realize the hierarchical input and ordered progressive calculation of the main factors and the secondary factors in the structure level; finally, according to the residual progressive calculation of each double residual block on the input data and the feature response summary, the contribution result of different environmental factor covariates to the biological quantity sequence is obtained and the prediction result is generated; based on this, in the whole technical scheme, the hierarchical analysis of the multi-layer residual structure can realize the layer-by-layer decoupling and contribution result quantification of the environmental factor influence, so as to obtain the prediction result of the change of plankton quantity under the current environment which is more structured and interpretable.

[0043] In one exemplary embodiment, the method further comprises step S301, and in addition, the method further comprises step S401.

[0044] Step S301, if the current double residual block represents the first double residual block in the double residual structure, the biological quantity sequence and the corresponding environmental factor covariate are taken as the input of the current double residual block, and the prediction value of the future residual and the fitting value of the biological quantity of the current double residual block are taken as the output of the current double residual block.

[0045] Step S401, if the current double residual block represents a double residual block other than the first double residual block in the double residual structure, the residual sequence between the biological quantity sequence and the fitting value output by the last double residual block, and the corresponding environmental factor covariate are taken as the input of the current double residual block, and the prediction value of the future residual and the fitting value of the biological quantity of the double residual block are taken as the output of the double residual block.

[0046] Wherein, the prediction value of the future residual represents the residual amount of the unexplained part in the future time axis inferred by the current double residual block in the calculation process according to the input signal, that is, the inference result of the unexplained part in the future time axis, further, it can be understood as the contribution result of the corresponding environmental factor covariate to the biological quantity sequence.

[0047] Wherein, the fitting value of the biological quantity represents the interpretable part calculated by the current double residual block in the calculation process, which is used to reflect the trend or periodic fitting result of the change of the biological quantity based on the corresponding environmental factor in the current layer.

[0048] The residual sequence represents a difference sequence between the biological quantity sequence and a fitting value output by the previous layer double residual block, and is used to represent a part of the biological quantity change that is not actually explained in the previous layer calculation process. Further, the residual sequence can be understood as residual information that is passed to the next layer.

[0049] For example, if the current double residual block represents the first double residual block in the double residual structure, the current double residual block belongs to the initial layer of double residual blocks. Thus, the biological quantity sequence and the corresponding environmental factor covariate are used as initial inputs to establish a starting benchmark for network calculation. Specifically, first, the input biological quantity sequence is time-index calibrated to ensure that the time steps of the input signal are consistent with the corresponding environmental factor covariate, thereby ensuring accurate pairing of the input signal in the time dimension. Subsequently, the two types of input signals are synchronously input into the current double residual block, and in the calculation process, the response characteristics of the residual sequence to the environmental factor are identified through linear combination and difference calculation, thereby distinguishing the explainable part from the unexplained part in the input signal. Furthermore, the current double residual block simultaneously establishes two output paths inside: one outputs a fitting value for the biological quantity, which represents the biological quantity change component that can be explained by the layer under the current input signal condition; and the other outputs a prediction value for the future residual, which represents the biological quantity change component that is not explained for the future period under the current input signal condition.

[0050] For example, if the current double residual block represents a double residual block other than the first double residual block in the double residual structure, the current double residual block does not belong to the initial layer of double residual blocks. Thus, the residual sequence between the biological quantity sequence and the fitting value output by the previous double residual block, and the corresponding environmental factor covariate are used as new input signals for supplementary calculation based on the unexplained part of the previous layer. Specifically, first, the biological quantity sequence and the fitting value output by the previous double residual block are subjected to time-step-by-time-step difference operation to obtain a residual sequence; the residual sequence reflects the fitting deviation of the biological quantity change by the previous layer, i.e., the dynamic component that is not explained by the previous layer. Subsequently, the residual sequence and the corresponding environmental factor covariate are respectively used as input signals of the current double residual block, so that the current layer can relearn the supplementary influence relationship of the environmental factor based on the residual of the previous layer; at the same time, the two types of input signals are time-index calibrated to ensure accurate pairing of the input signal in the time dimension. Subsequently, the two types of input signals are synchronously input into the current double residual block, and in the calculation process, the response characteristics of the residual sequence to the environmental factor are identified through linear combination and difference calculation, thereby further distinguishing the explainable part from the unexplained part in the input signal based on the unexplained part of the previous layer, and thus outputting a prediction value for the future residual and a fitting value for the biological quantity.

[0051] In this embodiment, firstly, the biological quantity sequence and the corresponding environmental factor covariate are taken as the input of the first double residual block, and the fitting value of the biological quantity and the prediction value of the future residual are output respectively, so as to realize the initial decomposition and trend modeling of the input signal, and provide the basic residual information for the subsequent layer; secondly, the residual sequence between the biological quantity sequence and the fitting value output by the last double residual block, and the corresponding environmental factor covariate are taken as the input, so as to perform supplementary calculation on the basis of the unexplained part of the previous layer, realize the layer-by-layer refinement and residual progressive correction of the environmental factor influence; based on this, in the whole technical scheme, the biological quantity change can be analyzed and fitted layer by layer through the multi-layer residual structure, so as to realize the high-precision progressive prediction of the plankton quantity change under the driving of multiple environmental factors.

[0052] In an exemplary embodiment, according to each double residual block, the contribution result of the corresponding environmental factor covariate to the biological quantity sequence is calculated respectively to obtain the prediction result of the plankton change amount under the current environment, including steps S501 to S504.

[0053] Step S501, in the current double residual block, the environmental factor covariate is subjected to feature extraction based on a parallel one-dimensional convolution network to obtain a covariate feature vector.

[0054] The parallel one-dimensional convolution network represents a neural computing structure that performs convolution operation in a sliding window manner in the time dimension, and is used to extract local time features and change trends from the environmental factor covariate.

[0055] Exemplarily, the input environmental factor covariate is subjected to feature extraction in the current double residual block to form a covariate feature vector that can express the change law of the time sequence. Specifically, firstly, the environmental factor covariate is subjected to feature extraction by a parallel one-dimensional convolution network, which contains multiple parallel convolution channels, each of which uses a convolution kernel of different length to capture the change mode of the environmental factor in different time scales. In the convolution calculation process, the sliding window mechanism is used to convolve step by step along the time direction to obtain the change trend of the environmental factor covariate in the local time range, and short-term fluctuation features and long-term change features are extracted in multiple parallel convolution channels. In addition, in order to avoid the scale difference between the features from causing deviation in calculation, the outputs of the parallel convolution channels are normalized and nonlinearly transformed to keep the feature amplitude distribution consistent.

[0056] Subsequently, the outputs of the parallel convolution channels are merged in the feature dimension to form a unified high-dimensional feature vector, i.e., a covariate feature vector, through channel concatenation and compression operations; the covariate feature vector contains multi-scale change information of the environmental factor in the time series in structure, i.e., becomes the intermediate output of the current double residual block, to support the subsequent associated modeling and residual progressive calculation process. In step S502, the covariate feature vector is calculated with another input received by the current double residual block to obtain a feature calculation result; wherein, if the current double residual block represents the first double residual block in the double residual structure, the other input received is the biological quantity sequence, 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 the residual sequence.

[0057] The feature cross calculation represents a process of weighted combination or mapping calculation of feature vectors from different input sources, for capturing the mutual dependence between different features.

[0058] The feature calculation result represents a composite feature expression formed after the feature cross calculation, for representing the associated structure between the environmental factor and the biological quantity change.

[0059] Exemplarily, first, the covariate feature vector extracted by convolution is received from the foregoing step, and another input signal corresponding to the hierarchical position of the current double residual block is selected: if the current double residual block is the initial layer in the structure, the other input signal is the original biological quantity sequence, for establishing the direct response relationship between the environmental factor and the biological quantity change; if the current double residual block is the subsequent layer in the structure, the other input signal is the residual sequence, for supplementary modeling based on the unexplained changes in the previous layer.

[0060] Subsequently, the covariate feature vector and the other input signal are cross-calculated at the same time index, i.e., the feature cross calculation result is generated by weighted combination or mapping calculation; for example, the calculation process realizes point-by-point fusion between the environmental factor related features and the biological quantity change related features in each time step, so that the dynamic change of the environmental factor corresponds to the biological quantity or residual information in the time series. Subsequently, the feature cross calculation result is normalized and combined to maintain the numerical continuity and scale consistency between time steps, thereby generating a structured feature calculation result.

[0061] Based on this, the feature calculation result numerically reflects the joint response strength of the environmental factor covariate to the biological quantity change in different time periods, and forms a composite feature set that can be used as the input of the internal calculation of the double residual block in structure.

[0062] Step S503, the feature calculation result is respectively calculated by two independent multi-layer perceptrons, and the prediction value of the current double residual block to the future residual and the fitting value to the biological quantity are obtained.

[0063] The multi-layer perceptron is a feedforward neural network structure composed of multi-layer linear mapping and nonlinear activation unit, which is used for nonlinear transformation and weighted calculation of the input feature calculation result, so as to output the corresponding prediction value or fitting value.

[0064] Exemplarily, in one aspect, the feature calculation result is input to one of the multi-layer perceptrons for processing. In the processing process, the feature calculation result is gradually transformed into high-dimensional feature representation in multi-layer mapping through the alternative operation of continuous linear transformation and nonlinear transformation, so as to extract the feature component in the input signal that can explain the change of biological quantity. Therefore, after multi-layer mapping, the multi-layer perceptron outputs the fitting value of the biological quantity, which is used to represent the part of the change of biological quantity that can be explained by the current double residual block according to the corresponding environmental factor.

[0065] On the other hand, the feature calculation result is input to the other multi-layer perceptron for processing. In the processing process, the feature calculation result is reconstructed in multi-layer calculation through the interlayer mapping of continuous linear transformation and nonlinear transformation, so as to extract the feature component that is not explained by the current layer structure. Therefore, after multi-layer mapping, the multi-layer perceptron outputs the prediction value of the future residual, which is used to represent the part of the change of biological quantity that cannot be explained by the current double residual block for the future period.

[0066] Based on this, the two multi-layer perceptrons are independent in parameters but consistent in input layer, so that the network can simultaneously generate the calculation results of biological quantity fitting and future residual prediction on the basis of the same input; therefore, such a double-path structure enables the network to output both explainable components and unexplained components for each double residual block, providing a continuous structural connection for subsequent residual transmission and accumulation calculation.

[0067] Step S504, combining the prediction values output by each double residual block, the prediction result of the change of plankton under the current environment is obtained.

[0068] Exemplarily, firstly, the prediction values output by each double residual block are respectively regarded as the contribution results of the corresponding environmental factor covariates to the biological quantity sequence, that is, each prediction value represents the unexplained component of the biological quantity change in the corresponding double residual block for the future time period with respect to the corresponding environmental factor. Subsequently, the contribution results are aggregated and weighted integrated in the time dimension to form a predicted overall residual result reflecting the overall residual change in the future time period; wherein the predicted overall residual result contains the cumulative effects of all environmental factors in the multi-layer residual structure, reflecting both long-term trends and local fluctuations. Then, on the basis of the original biological quantity sequence, the residual signals in the predicted overall residual result are supplemented to the subsequent positions of the original biological quantity sequence in chronological order, for example, the value of the original biological quantity sequence at the last time point can be taken as the basis value and the predicted value is superimposed, thereby finally outputting the biological quantity prediction result evolving over time, so that the historical observation data can be extended to the future in the time dimension, thereby realizing the continuous connection and smooth transition from historical quantity change to future quantity prediction.

[0069] Based on this, the prediction sequence with time coherence and causal consistency is generated, that is, the prediction result based on the change amount of plankton under the current environment, thereby embodying the cumulative influence and dynamic coupling characteristics of different environmental factors on the biological quantity change under the current environmental conditions, completing the mapping process from hierarchical residual calculation to overall trend reconstruction.

[0070] Optionally, when the contribution results output by each double residual block are weighted integrated in the time dimension, the weights can be allocated according to the hierarchical position of each double residual block in the residual progressive calculation and its explanatory degree to the biological quantity sequence. That is, the contribution results at the front level represent the main trend influence, and the weight is higher; the contribution results at the back level reflect the subtle or compensatory influence, and the weight is relatively lower; thereby, in the process of weighted integration, the proportion of each contribution result in the overall residual fusion is controlled by the weight proportion between layers, so that the deep residual remains structural integrity while not weakening the main residual contribution.

[0071] In this embodiment, first, according to the feature extraction of the environmental factor covariate through the parallel one-dimensional convolution network, a multi-scale covariate feature vector is obtained in the time dimension, providing a structured input for subsequent cross calculation; second, according to the feature cross calculation of the covariate feature vector and another input of the current double residual block to obtain a feature calculation result, thereby establishing the relationship between the environmental factor and the change of the biological quantity, realizing the corresponding fusion between different input signals; then, according to the input of the feature calculation result into two multilayer perceptrons for independent calculation, the fitting value of the biological quantity and the prediction value of the future residual are obtained respectively, realizing the synchronous modeling of the current layer to the explainable part and the unexplained part; based on this, in the whole technical scheme, through the collaborative design of multi-layer residual structure and multi-stage feature calculation, the hierarchical prediction and progressive mapping of the change of the biological quantity under the action of multiple environmental factors are realized.

[0072] In an exemplary embodiment, the covariate feature vector and another input received by the current double residual block are cross calculated to obtain a feature calculation result, including steps S601 to S603.

[0073] In step S601, according to a preset learning mechanism based on hidden unit response, the covariate feature vector is nonlinearly transformed to obtain a scaling factor sequence for adjusting the hidden unit response.

[0074] Wherein, the hidden unit represents an internal calculation node for feature transformation in the network structure, which is used for weighted summation and nonlinear mapping of input features; the hidden unit response represents the response output obtained after the hidden unit processes the input features, which is used to represent the response intensity of the activated input features in the hidden unit.

[0075] Wherein, the learning mechanism based on the hidden unit response represents a computer mechanism that uses the activation output features of the hidden unit for parameter adaptive update, which is used to adjust the weight distribution inside the network according to the response output corresponding to different input features.

[0076] Exemplarily, first, the covariate feature vector is input into the hidden unit inside the double residual block, and each hidden unit respectively performs weighted summation and nonlinear mapping after receiving the input features, thereby generating the response output corresponding to each hidden unit respectively. Further, in the learning mechanism based on the hidden unit response, according to the distribution characteristics of the hidden unit response output, the activity degree of the hidden unit in different time steps and feature dimensions is counted, and the internal parameters are adjusted accordingly, so that the network can adaptively identify the sensitive area of the input features; wherein, the hidden unit with stronger response is given higher weight, and the hidden unit with weaker response is relatively inhibited, to form a dynamic response structure for the difference of the input features.

[0077] Based on this, according to the above adaptive updating process, the importance of each input feature in the time sequence can be gradually learned, thereby forming a scaling factor sequence at the output end; wherein the numerical value of the scaling factor sequence reflects the response weight of different inputs at a certain time step, which acts on the input features through the learning mechanism for proportional adjustment, thereby serving as the intermediate output of the current double residual block, which not only retains the dynamic change information of the input features, but also completes the adaptive learning of the feature weight.

[0078] In step S602, another input received by the current double residual block is element-wise weighted according to the scaling factor sequence to obtain a weighted result.

[0079] Exemplarily, first, the scaling factor sequence is synchronized with the time index of the biological quantity sequence or the residual sequence received by the current double residual block, so that they maintain a one-to-one correspondence in the time step. Subsequently, in each time step, the corresponding scaling factor in the scaling factor sequence is multiplied element-wise with the numerical value of the input signal, so that the amplitude of the input signal is proportionally adjusted according to the size of the scaling factor, thereby reflecting the difference of the input features in the time sequence in the numerical level; that is, after the scaling factor sequence is trained by the learning mechanism in the foregoing step, it can reflect the importance and sensitivity of the input signal at each time step, therefore, through the element-wise weighting operation, the dynamic control of the input response strength can be realized without changing the time structure of the input signal.

[0080] With the progression of the time step, the input signal is continuously weighted to form a weighted sequence that has been adjusted in value by the feature sensitivity; in addition, in order to maintain the balance of the overall data, normalization and scale constraint processing are performed on the result after weighting calculation to prevent excessive deviation due to too large scaling factor difference in different time periods.

[0081] Finally, the result after weighting and calibration is defined as the weighted result, which reflects the distribution adjustment of the input signal in the time dimension in the numerical level, preserves the time order and correlation of the original signal in the structural level, and reflects the dynamic adjustment effect of the environmental factor on the input signal in the semantic level.

[0082] In step S603, the weighted result is combined with the covariate feature vector to obtain a feature calculation result.

[0083] Exemplarily, the weighted result is combined with the covariate feature vector in a linear superposition manner, for example, point-by-point fusion between the biological quantity change related features and the environmental factor related features is realized in each time step, so that the two types of features embodied in the weighted result are fused in the same space. That is, the weighted result reflects the dynamic change of the input signal under the action of the scaling factor sequence, and the covariate feature vector retains the global feature distribution of the environmental factor, and the combination of the two can express both time dependence and overall environmental characteristics, thereby enhancing the explanation ability of the input signal. Based on this, after the fusion processing, the obtained result is taken as a feature calculation result, which not only contains the scaled and adjusted change information of the biological quantity, but also integrates the feature expression of the environmental factor in structure, thereby forming a composite feature expression that can be used for progressive calculation of the multi-layer residual structure.

[0084] In the embodiment, first, a learning mechanism based on hidden unit response is implemented on the covariate feature vector to generate a scaling factor sequence, so that the network can adaptively identify the importance of the input features in the time dimension and realize dynamic adjustment of the response weight; second, the biological quantity sequence or the residual sequence is weighted element by element according to the scaling factor sequence, so that it reflects the response intensity and influence difference of different input features in the time sequence; third, the weighted result is fused with the covariate feature vector to form a composite feature representation that reflects both the biological quantity change and the environmental factor influence; based on this, in the whole technical solution, through the learning mechanism based on the hidden unit response, adaptive mapping and unified representation between the environmental factor related features and the biological quantity change are realized.

[0085] In an exemplary embodiment, the covariate feature vector is calculated with another input received by the current double residual block to obtain a feature calculation result, including steps S701 to S702.

[0086] Step S701, first-order difference calculation is performed on the biological quantity sequence to obtain a biological change rate sequence.

[0087] The biological change rate sequence represents a time sequence for describing the change speed of the biological quantity between adjacent time steps, for example, reflecting the growth or reduction amplitude of a certain type of plankton in a continuous monitoring period.

[0088] Exemplarily, firstly, for the biological quantity sequence, a point-by-point difference operation is performed between adjacent time steps to obtain a difference result, i.e., the observation value at the next time minus the observation value at the previous time, so as to obtain the change difference between each time step. Subsequently, the time index alignment and boundary smoothing processing are performed on the difference result to ensure that the obtained biological change rate sequence is consistent with the original biological quantity sequence in the time dimension. Based on this, the biological change rate sequence reflects the change speed of the biological quantity in adjacent time periods, and further, the first-order difference processing process can eliminate the cumulative trend effect in the original biological quantity sequence, so that the input data can reflect the dynamic change characteristics without being affected by the total level.

[0089] In step S702, the covariate feature vector is cross-calculated with another input received by the current double residual block to obtain a feature calculation result; wherein if the current double residual block represents the first double residual block in the double residual structure, the other input received is the biological change rate sequence, 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 the target residual sequence between the biological change rate sequence and the fitting value output by the previous double residual block.

[0090] Exemplarily, firstly, the covariate feature vector extracted by convolution is received from the foregoing step, and another input signal corresponding to the hierarchical position of the current double residual block is selected according to the hierarchical position of the current double residual block: if the current double residual block is the initial layer in the structure, the other input signal is the biological change rate sequence, which is used to establish the direct response relationship between the environmental factor and the biological quantity change speed; if the current double residual block is a subsequent layer in the structure, the other input signal is replaced by the target residual sequence between the biological change rate sequence and the fitting value output by the previous double residual block, which is used to supplement modeling based on the change not explained by the previous layer.

[0091] Subsequently, the covariate feature vector and the other input signal are cross-calculated at the same time index, i.e., the cross-calculation is realized by corresponding calculation at each time step, and the covariate feature vector and the other input signal are weighted and combined at the numerical level to form a composite feature representation describing the joint influence of the environmental factor and the biological quantity change rate. As can be seen, the core of this calculation process is to retain the global trend information of the environmental factor, while superimposing the local dynamic change reflected by the biological quantity change rate, thereby generating a feature calculation result realizing synchronous modeling of environmental action and biological dynamic change.

[0092] In this embodiment, first, the original quantity information is converted into a biological change rate sequence representing the change rate of adjacent time according to the first-order difference calculation on the biological quantity sequence, thereby providing more sensitive change characteristics for subsequent calculation; second, the feature cross calculation is performed in the double residual block according to another input related to the covariate feature vector and the biological change rate sequence, thereby establishing the dynamic mapping relationship between the environmental factors and the biological change rate, and achieving the supplementary modeling of unexplained changes through hierarchical structure progression; based on this, in the whole technical solution, the network's analysis and prediction ability for the change process of phytoplankton quantity is enhanced on both time and feature levels.

[0093] In one exemplary embodiment, Figure 2 An N-BEATSx network prediction method fusing multi-source environmental factors is shown, specifically: first, on the one hand, the phytoplankton quantity sequence is obtained by first-order difference operation to obtain a biological change rate sequence; on the other hand, N environmental factor covariates are obtained by covariate analysis on multi-source environmental data, each environmental factor covariate corresponding to a type of environmental factor.

[0094] Secondly, the long-term correlation between each environmental factor covariate and the phytoplankton quantity sequence or the biological change rate sequence is analyzed and sorted, so as to distribute each environmental factor covariate to the corresponding double residual block according to the strength of the long-term correlation; for example, the environmental factor covariate with the strongest long-term correlation is taken as the input of the first double residual block STACK1, and the environmental factor covariate with the weakest long-term correlation is taken as the input of the last double residual block STACKn; wherein the number of double residual blocks matches the number of environmental factor covariates.

[0095] Secondly, the biological change rate sequence and the environmental factor covariate with the first long-term correlation ranking are taken as the input of the first double residual block STACK1, and STACK1 is processed accordingly to obtain a prediction residual sequence and a regression sequence; wherein the prediction residual sequence represents the prediction value of future residual in the sequence level, and the regression sequence represents the fitting value of the biological quantity in the sequence level.

[0096] Secondly, the regression sequence output by STACK1 and the biological change rate sequence are subjected to difference calculation to obtain a residual sequence, and the residual sequence and the environmental factor covariate with the second long-term correlation ranking are taken as the input of the second double residual block STACK2, and STACK2 is processed accordingly to obtain a prediction residual sequence and a regression sequence.

[0097] Further, the recursive residual calculation is performed on the N double residual blocks to obtain the prediction residual sequence output by each double residual block respectively; each prediction residual sequence is taken as the contribution result of the biological change rate sequence of each environmental factor covariate, that is, the observation value at the last time point of the biological change rate sequence is superimposed with each prediction residual sequence, thereby obtaining the prediction result of the plankton change amount under the current environment output by the improved N-BEATSx network.

[0098] In an exemplary embodiment, Figure 3 The structure diagram of a double residual block of an improved N-BEATSx network is shown, specifically: first, the biological change rate sequence is calculated with the regression sequence output by the last double residual block STACK(i-1) to obtain a residual sequence, and the residual sequence and the i-th environmental factor covariate in the long-term correlation order are taken as the input of the i-th double residual block STACKi.

[0099] On the one hand, the residual sequence is linearly mapped and nonlinearly transformed through a fully connected layer (FCN, Fully Connected Network) to extract the time feature and trend feature of the residual sequence, and to convert the original input into a feature representation that can be aligned with the environmental feature; on the other hand, the environmental factor covariate is sequentially convolved through a time convolution network (TCN, Temporal Convolutional Network) to capture its local change rule and periodicity feature in the time dimension, and outputs a covariate feature vector.

[0100] Further, the output features of the fully connected layer and the covariate feature vector are multiplied by Hadamard product, i.e. element-wise multiplication, to realize the interactive fusion of the two features, so that the network can adaptively adjust the response strength of the residual information to the environmental factor at each time step.

[0101] Further, the feature components of the features obtained by the Hadamard product operation are decomposed in multiple stages through a progressive calculation unit (M*blocks) composed of M residual sub-modules, so that the network gradually approximates the true structure of the residual distribution; each block contains a forward prediction and error compensation path to ensure deep feature learning ability within the layer.

[0102] Further, on the one hand, the output of the M*blocks is linearly mapped through a prediction linear layer to generate the prediction residual sequence of the current layer; on the other hand, the output of the M*blocks is independently regressed through a regression linear layer to generate the regression sequence of the current layer.

[0103] It should be understood that although the steps in the flowcharts involved in the embodiments described above are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the embodiments described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or steps or stages in other steps.

[0104] Based on the same inventive concept, the embodiments of the present application also provide a fusion multi-source environmental factor N-BEATSx network prediction device for implementing the fusion multi-source environmental factor N-BEATSx network prediction method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more fusion multi-source environmental factor N-BEATSx network prediction device embodiments provided below can refer to the limitations of the fusion multi-source environmental factor N-BEATSx network prediction method described above, and will not be repeated here.

[0105] In one exemplary embodiment, as shown in Figure 4 A fusion multi-source environmental factor N-BEATSx network prediction device is provided, comprising: an acquisition module 101, a conversion module 102, and an analysis module 103, wherein: The acquisition module 101 is configured to acquire multi-source environmental data and plankton observation data in a current environment; The conversion module 102 is configured to convert the multi-source environmental data into different environmental factor covariates, and convert the plankton observation data into a biological quantity sequence; The analysis module 103 is configured to input each environmental factor covariate and the biological quantity sequence into a preset improved N-BEATSx network, calculate the contribution of each environmental factor covariate to the biological quantity sequence according to the improved N-BEATSx network, and obtain a prediction result based on the plankton change amount in the current environment.

[0106] In an example embodiment, the analysis module 103 is further configured to: calculate long-term correlations between each environmental factor covariate and the biological quantity sequence respectively, and obtain a ranking result of each environmental factor covariate according to the corresponding long-term correlation; determine, according to a preset double residual structure in the improved N-BEATSx network, a corresponding double residual block of each environmental factor covariate in the double residual structure combined with the ranking result; and calculate, according to each double residual block, a contribution result of the corresponding environmental factor covariate to the biological quantity sequence respectively, to obtain the prediction result based on the change amount of plankton under the current environment.

[0107] In an example embodiment, the analysis module 103 is further configured to: if the current double residual block represents the first double residual block in the double residual structure, take the biological quantity sequence and the corresponding environmental factor covariate as inputs of the current double residual block respectively, and take the predicted value of the future residual and the fitting value of the biological quantity of the current double residual block as outputs of the current double residual block respectively.

[0108] In an example embodiment, the analysis module 103 is further configured to: if the current double residual block represents a double residual block other than the first double residual block in the double residual structure, take the residual sequence between the biological quantity sequence and the fitting value output by the last double residual block, and the corresponding environmental factor covariate as inputs of the current double residual block respectively, and take the predicted value of the future residual and the fitting value of the biological quantity of the double residual block as outputs of the double residual block respectively.

[0109] In an example embodiment, the analysis module 103 is further configured to: in the current double residual block, perform feature extraction on the environmental factor covariate based on a parallel one-dimensional convolutional network to obtain a covariate feature vector; and perform feature cross calculation on the covariate feature vector and another input received by the current double residual block to obtain a feature calculation result; wherein, if the current double residual block represents the first double residual block in the double residual structure, the other input received is the biological quantity sequence, 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 the residual sequence; and perform output calculation on the feature calculation result by two independent multilayer perceptrons respectively to obtain the predicted value of the future residual and the fitting value of the biological quantity of the current double residual block; and combine the predicted values output by each double residual block to obtain the prediction result based on the change amount of plankton under the current environment.

[0110] In an example embodiment, the analysis module 103 is further configured to perform a nonlinear transformation on the covariate feature vector according to a preset learning mechanism based on the hidden unit response to obtain a sequence of scaling factors for adjusting the hidden unit response; perform an element-wise weighting processing on another input received by the current double residual block according to the sequence of scaling factors to obtain a weighting result; and combine the weighting result with the covariate feature vector to obtain the feature calculation result.

[0111] In an example embodiment, the analysis module 103 is further configured to perform a first-order difference calculation on the biological quantity sequence to obtain a biological change rate sequence; and perform a feature cross calculation on the covariate feature vector and another input received by the current double residual block to obtain the feature calculation result; wherein, if the current double residual block represents a first double residual block in the double residual structure, the another input received is the biological change rate sequence, 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 another input received is a target residual sequence between the biological change rate sequence and a fitting value of an output of a previous double residual block.

[0112] The above-mentioned modules in the N-BEATSx network prediction device fusing multi-source environmental factors can be realized by software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to the above-mentioned modules.

[0113] In an example embodiment, a computer device is provided, which includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above-mentioned embodiments.

[0114] In an embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps in any of the above-mentioned embodiments.

[0115] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments.

[0116] The technical features of the above-mentioned embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0117] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A prediction method for N-BEATSx networks that integrates multiple environmental factors, characterized in that, The method includes: Acquire multi-source environmental data and plankton observation data under the current environment; The multi-source environmental data is converted into different environmental factor covariates, and the plankton observation data is converted into a biological quantity sequence; Each environmental factor covariate and the biomass sequence are input into a preset improved N-BEATSx network. The contribution of each environmental factor covariate to the biomass sequence is calculated based on the improved N-BEATSx network to obtain the prediction result based on the change in plankton under the current environment.

2. The method according to claim 1, characterized in that, The step of calculating the contribution of the environmental factor covariates to the biological abundance sequence based on the improved N-BEATSx network to obtain the prediction results based on the change in plankton under the current environment includes: The long-term correlation between each environmental factor covariate and the biological quantity sequence is calculated to obtain the ranking results of each environmental factor covariate according to the corresponding long-term correlation. Based on the pre-defined double residual structure in the improved N-BEATSx network, and combined with the ranking results, determine the double residual blocks corresponding to each environmental factor covariate in the double residual structure. Based on each double residual block, the contribution of the corresponding environmental factor covariates to the biological quantity sequence is calculated to obtain the prediction results based on the change in plankton quantity under the current environment.

3. The method according to claim 2, characterized in that, The method further includes: If the current double residual block represents the first double residual block in the double residual structure, then the biological quantity sequence and the corresponding environmental factor covariates are used as the inputs of the current double residual block, and the predicted value of the current double residual block for the future residual and the fitted value for the biological quantity are used as the outputs of the current double residual block.

4. The method according to claim 2, characterized in that, The method further includes: 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 biological quantity sequence and the fitted value output by the previous double residual block, as well as the corresponding environmental factor covariates, are used as the inputs of the current double residual block, and the predicted value of the double residual block for future residuals and the fitted value for biological quantity are used as the outputs of the double residual block.

5. The method according to claim 3 or 4, characterized in that, The step of calculating the contribution of corresponding environmental factor covariates to the biological abundance sequence based on each double residual block to obtain the prediction result based on the change in plankton under the current environment includes: In the current dual residual block, feature extraction of the environmental factor covariates is performed based on a parallel one-dimensional convolutional network to obtain the covariate feature vector; The covariate feature vector is subjected to feature cross-computation with the other input received by the current double residual block to obtain the feature calculation result; wherein, if the current double residual block represents the first double residual block in the double residual structure, the other input received is the biological quantity sequence; 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. The feature calculation results are output by two independent multilayer perceptrons to obtain the predicted value of the current double residual block for the future residual and the fitted value for the number of organisms. By combining the predicted values ​​output by each dual residual block, the prediction result based on the change in plankton in the current environment is obtained.

6. The method according to claim 5, characterized in that, The step of performing feature cross calculation between the covariate feature vector and another input received by the current dual residual block to obtain the feature calculation result includes: According to a preset learning mechanism based on hidden unit response, the covariate feature vector is nonlinearly transformed to obtain a scaling factor sequence for adjusting the hidden unit response. The other input received by the current double residual block is weighted element-wise according to the scaling factor sequence to obtain a weighted result; The weighted result is combined with the covariate feature vector to obtain the feature calculation result.

7. The method according to claim 5, characterized in that, The step of performing feature cross calculation between the covariate feature vector and another input received by the current dual residual block to obtain the feature calculation result includes: The biological quantity sequence is subjected to first-order difference calculation to obtain the biological change rate sequence; The covariate feature vector is subjected to feature cross-computation with the other input received by the current double residual block to obtain the feature calculation result; wherein, if the current double residual block represents the first double residual block in the double residual structure, the other input received is the biological rate of change sequence; 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 target residual sequence between the biological rate of change sequence and the fitted value output by the previous double residual block.

8. An N-BEATSx network prediction device that integrates multiple environmental factors, characterized in that, The device includes: The acquisition module is used to acquire multi-source environmental data and plankton observation data under the current environment; The conversion module is used to convert the multi-source environmental data into different environmental factor covariates and to convert the plankton observation data into a biological quantity sequence. The analysis module is used to input various environmental factor covariates and the biomass sequence into a preset improved N-BEATSx network, and calculate the contribution of each environmental factor covariate to the biomass sequence based on the improved N-BEATSx network to obtain the prediction result based on the change in plankton under the current environment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images