Fourier self-adaptive variable fusion time series prediction method and device

Abstract

The embodiment of the application provides a variable fusion time series prediction method and device based on Fourier self-adaption and a storage medium, the method comprising: acquiring a population quantity observation time series of a target pest in a historical time period, and acquiring a multi-dimensional meteorological covariant time series synchronous with the historical time period; performing cross-correlation variable fusion on the population quantity observation time series and the multi-dimensional meteorological covariant time series to generate a fusion feature sequence; performing frequency domain transformation on the fusion feature sequence, extracting at least one frequency component corresponding to a core physical period whose energy intensity satisfies a preset condition to form a period set; performing dynamic resampling on the fusion feature sequence according to the period set to generate a plurality of adaptive block matrices; performing double-weighted fusion on the plurality of adaptive block matrices to output a comprehensive feature tensor; and performing prediction according to the comprehensive feature tensor to output a future pest population size prediction sequence.

Description

Technical Field

[0001] This application relates to the field of agricultural pest prediction technology, specifically to a method, device, and storage medium for time series prediction based on Fourier adaptive variable fusion. Background Technology

[0002] Predicting the population dynamics of agricultural pests is a key technological link in achieving early warning and scientific control of crop diseases and pests. Major rice pests, represented by the rice stem borer (Chilo suppressalis), exhibit typical periodicity in their population outbreaks. These outbreaks are driven by the endogenous rhythms of their own generational succession and significantly influenced by external meteorological conditions such as temperature, humidity, and precipitation, making it a typical multivariate time series prediction problem.

[0003] Existing time series forecasting methods for agricultural pest forecasting primarily employ direct concatenation or equal-weighted fusion of pest population sequences with meteorological covariates as input into the forecasting model. However, meteorological sequences contain a significant amount of random noise with low correlation to pest dynamics. This indiscriminate fusion method leads to redundant information interfering with the core evolutionary trajectory of pest reproduction, making it difficult for the model to focus on the essential patterns of the target variable. Furthermore, existing methods generally employ fixed-length time series block strategies, forcibly truncating the continuous biological evolutionary process of pests. This fragments the complete reproductive cycle across multiple data blocks, causing semantic fragmentation and phase drift. Simultaneously, the equal-weighting of features across different time periods in existing models fails to highlight the informational value of key precursor windows preceding pest outbreaks, diluting and obscuring weak but crucial surge signals. These shortcomings limit the accuracy and reliability of pest population forecasting. Summary of the Invention

[0004] The purpose of this application is to provide a Fourier adaptive variable fusion time series prediction method, device and storage medium to solve the technical problems in the prior art, such as severe meteorological noise interference, fixed block strategy leading to the fragmentation of biological evolution cycle, and feature equalization processing leading to the dilution and submergence of key precursor signals before pest outbreaks.

[0005] To achieve the above objectives, this application provides an embodiment of a variable fusion time series prediction method based on Fourier adaptive variability fusion, the method comprising: Obtain the observation time series of the population size of the target pest within a historical time period, and obtain the time series of multidimensional meteorological covariates for the same period in history; Cross-correlation variables were fused between the time series of population size observations and the time series of multidimensional meteorological covariates to generate a fused feature sequence. A frequency domain transformation is performed on the fused feature sequence to extract the core physical period corresponding to at least one frequency component whose energy intensity meets the preset conditions, thus forming a period set. Based on the period set, the fused feature sequence is dynamically resampled to generate multiple sets of adaptive block matrices; each set of adaptive block matrices corresponds to a core physical period in the period set, and the block scale in each set of adaptive block matrices is associated with the core physical period corresponding to that set. Perform double-weighted fusion on multiple sets of adaptive block matrices to output a comprehensive feature tensor; Based on the comprehensive feature tensor, a prediction sequence for the future pest population size is output.

[0006] In this embodiment, the population size observation time series and the multidimensional meteorological covariate time series are cross-correlated variables fused to generate a fused feature sequence, including: performing a fast Fourier transform on the fused feature sequence to calculate the amplitude value of each frequency component; setting an effective period interval according to the biological reproduction characteristics of the target pest and filtering out frequency components outside the effective period interval; sorting each frequency component retained in the effective period interval according to the amplitude value from largest to smallest; determining the first K frequency components after sorting as selected frequency components, where K is a preset positive integer; and extracting the period corresponding to each selected frequency component to form a period set.

[0007] In this embodiment, the fused feature sequence is dynamically resampled according to the period set to generate multiple sets of adaptive block matrices. This includes: calculating the ideal stretch length for each core physical period in the period set; using a linear interpolation algorithm to resample the fused feature sequence to the ideal stretch length to generate an interpolation sequence; dividing the interpolation sequence into a preset number of non-overlapping blocks, where the physical time sequence length of each block is equal to the currently processed core physical period; and reshaping the dimensions of all the equally divided blocks to generate a set of adaptive block matrices corresponding to the core physical period.

[0008] In this embodiment of the application, for each core physical cycle in the cycle set, the ideal stretch length is calculated, including: obtaining the preset total number of blocks, the cycle duration of the current core physical cycle, and the original time sequence length of the fused feature sequence; taking the product of the preset total number of blocks and the cycle duration of the current core physical cycle as the theoretical sequence length; comparing the numerical values ​​of the theoretical sequence length and the original time sequence length of the fused feature sequence, and selecting the larger of the two as the ideal stretch length corresponding to the current core physical cycle.

[0009] In this embodiment, prediction based on a comprehensive feature tensor to output a future pest population size prediction sequence includes: flattening the comprehensive feature tensor along all dimensions to obtain a one-dimensional global context vector; inputting the one-dimensional global context vector into a multi-layer fully connected network, which contains multiple fully connected layers connected sequentially, with adjacent fully connected layers undergoing nonlinear transformation through Gaussian error linear units; directly mapping the one-dimensional global context vector through the multi-layer fully connected network to convert it into an output vector with a dimension equal to the target prediction horizon step size; and using the output vector as the future pest population size prediction sequence within a preset time period.

[0010] In this embodiment, the process of fusing cross-correlation variables between the population size observation time series and the multidimensional meteorological covariate time series to generate a fused feature sequence further includes: concatenating and splicing the population size observation time series and the multidimensional meteorological covariate time series along the feature dimension to obtain a joint matrix; performing a one-dimensional convolution operation on the joint matrix along the time dimension to extract the co-evolution features between endogenous and exogenous variables and generate a cross-correlation auxiliary feature representation; and using learnable scalar parameters, with the population size observation time series as the dominant factor, weightedly adding the population size observation time series and the cross-correlation auxiliary feature representation to output the fused feature sequence.

[0011] In this embodiment, a learnable scalar parameter is used, with the population size observation time series as the dominant sequence. The population size observation time series and the cross-correlation auxiliary feature representation are weighted and their residuals are summed to output a fused feature sequence. This includes: using the population size observation time series as the dominant sequence and the cross-correlation auxiliary feature representation as auxiliary information; initializing the learnable scalar parameter, which controls the information ratio between the population size observation time series and the cross-correlation auxiliary feature representation during the fusion process; calculating the product of the learnable scalar parameter and the population size observation time series to obtain the dominant weighted feature; calculating the product of the cross-correlation auxiliary feature representation and the target coefficient to obtain the auxiliary weighted feature, where the target coefficient is the difference between 1 and the learnable scalar parameter; and summing the dominant weighted feature and the auxiliary weighted feature to obtain the fused feature sequence.

[0012] In this embodiment, a dual-weighted fusion is performed on multiple sets of adaptive block matrices to output a comprehensive feature tensor. This includes: performing adaptive average pooling on the interpolation sequence corresponding to the current scale, compressing the temporal length of the interpolation sequence back to the original length of the fused feature sequence to obtain a pooled sequence; flattening the pooled sequence into a one-dimensional vector and inputting it into a linear evaluation network to output an attention weight vector with a dimension equal to the preset total number of blocks, and then normalizing the attention weight vector using a normalized exponential function; multiplying each normalized weight scalar into each temporal block to complete the focusing of time-dimension differentiated features; performing adaptive average pooling again on all weighted temporal blocks to uniformly compress all blocks to a fixed standard block length and outputting a standardized feature matrix; and performing equal-weighted superposition fusion on the standardized feature matrices obtained under different core physical cycle scales to obtain a comprehensive feature tensor.

[0013] A second aspect of this application provides a Fourier adaptive variable fusion time series prediction device, the device comprising: Memory, used to store computer programs; A processor for executing a computer program to implement any of the above-mentioned Fourier adaptive variable fusion time series forecasting methods.

[0014] A third aspect of this application provides a machine-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform any of the above-described Fourier adaptive variable fusion time series prediction methods.

[0015] According to the above technical solution, this application, through a cross-correlation variable fusion module, takes the time series of target pest population observations as the main factor and injects multidimensional meteorological covariates as auxiliary information, rather than weighting them equally or directly splicing them together. This effectively suppresses the interference of redundant random noise in the meteorological sequence on the main line of pest reproduction and evolution, allowing the model to focus on the essential laws of the target variable and improve the accuracy of prediction. By performing frequency domain transformation on the fused feature sequence, the core physical cycle with the strongest energy is adaptively extracted from the data. Based on the extracted cycle, the fused feature sequence is dynamically resampled to generate an adaptive block matrix, so that the block scale is associated with the real biological rhythm cycle of the pest population. This avoids the problem of the biological evolution process being forcibly truncated and the complete reproduction cycle being fragmented and scattered in multiple data blocks due to fixed-length blocks, thus preserving the natural continuity and semantic integrity of the time series information to the greatest extent. By performing dual-weighted fusion on multiple sets of adaptive block matrices, and using a block-level attention mechanism to differentiate the weights of each block in the time dimension, the model accurately locks the key precursor window before pest outbreaks, preventing weak but crucial local surge signals from being diluted and submerged. At the same time, by using multi-scale equal-weighted integration to synthesize evolutionary features under different physical cycle scales, the model takes into account the advantages of both long and short cycles at the macro level, improving its generalization ability and robustness in complex meteorological environments.

[0016] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A flowchart illustrating a Fourier adaptive variable fusion time series prediction method according to an embodiment of this application is shown in the schematic diagram. Figure 2 A schematic diagram of a data preprocessing flow according to an embodiment of this application is shown. Detailed Implementation

[0018] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the embodiments of this application.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] In one embodiment of this application, as Figure 1 As shown, a variable fusion time series prediction method based on Fourier adaptive spectroscopy is provided, the steps of which include: Step 101: Obtain the population size observation time series of the target pest within the historical time period, and obtain the multidimensional meteorological covariate time series of the same period as the historical time period.

[0021] In one embodiment of this application, the target pest is the rice stem borer. The method for obtaining the population size observation time series can be to periodically collect pest population size data in the target monitoring area according to a unified and standardized survey method. The collection methods include, but are not limited to, one or more of the following: light trap monitoring, sex pheromone trap monitoring, field system survey, and automatic counting by intelligent insect monitoring lamps. The data obtained by different collection methods are arranged in chronological order to form a population size observation time series with daily time units.

[0022] Taking the intelligent insect monitoring lamp as an example, the lamp collects and automatically counts the insects it traps daily, generating a daily population time series. For missing data due to equipment malfunction, network interruption, or severe weather, it is supplemented in subsequent preprocessing steps.

[0023] In this embodiment of the application, the time series of population size observations can be represented as: Where L is the total length of the historical time window, This represents the number of pest populations observed in the t-th time unit, where t is the time index.

[0024] The multidimensional meteorological covariate time series can be obtained by acquiring meteorological observation data consistent with the population size observation time series from agricultural meteorological monitoring stations or meteorological departments within the target monitoring area. Meteorological covariates include, but are not limited to, multiple parameters such as average temperature, maximum temperature, minimum temperature, relative humidity, precipitation, sunshine duration, evaporation, wind speed, and ground temperature. The selection of each meteorological factor in the multidimensional meteorological covariate time series is determined based on the biological characteristics of the target pest.

[0025] For the rice stem borer, five core meteorological factors were selected: average temperature, daily maximum temperature, daily minimum temperature, daily sunshine duration, and daily evaporation. Among these, average daily temperature, daily maximum temperature, and daily minimum temperature collectively influence the development rate and generation cycle of the rice stem borer. The generation cycle refers to the total number of days required for the rice stem borer to complete a full life cycle (from egg to adult). Daily sunshine duration affects the induction of diapause in the rice stem borer; when sunshine duration decreases to a certain threshold, the larvae enter diapause to overwinter. Daily evaporation affects the field microclimate, thus indirectly affecting the survival and development of larvae. These five meteorological factors have a significant biological correlation with the population dynamics of the rice stem borer.

[0026] In this embodiment of the application, the time series of multidimensional meteorological covariates can be represented as follows: ,in, Let be a multidimensional meteorological covariate vector at the t-th time unit, where each vector contains multiple meteorological factor dimensions, i.e.: , where d is the total number of dimensions of the meteorological covariates.

[0027] The time series of population size observations and the time series of multidimensional meteorological covariates are strictly aligned on the time axis to ensure that the same time index t corresponds to the same monitoring date, thus ensuring the consistency of temporal semantics when subsequent variables are fused.

[0028] In the embodiments of this application, the historical time window length L of the population size observation time series is dynamically determined according to the biological characteristics of the target pest, and is generally not less than one complete pest generation cycle to ensure that the sequence contains sufficient periodic information. For pests with short generation cycles, such as aphids, the time window length can be 90 to 180 days; for pests with long generation cycles, such as rice stem borers, the time window length is preferably 365 to 730 days.

[0029] It should be noted that, in the embodiments of this application, the multidimensional meteorological covariate time series and the population size observation time series are fully aligned on the time axis, that is, the two have the same time start, time end and time interval, to ensure that the population size data at each time point and the corresponding meteorological data are consistent in temporal semantics during subsequent cascading splicing and variable fusion.

[0030] In one embodiment of this application, as Figure 2 As shown, it also includes a data preprocessing step. Specifically, during the data acquisition process, due to factors such as inconsistent observations, sensor malfunctions, and human reporting errors, missing values ​​are unavoidable in the population size observation time series and the meteorological covariate time series. The processor uses linear interpolation to complete the missing values, estimating the missing data while preserving the underlying trend of the data. For the missing data point at position k, the interpolation calculation formula is: ,in, and Let be the values ​​of the two nearest known data points before and after the missing position, and let a and b be the time positions of these two known data points, respectively. This method assumes a local linear relationship between adjacent observations, ensuring a seamless transition of the reconstructed time series without introducing additional bias or abrupt changes.

[0031] Furthermore, during the collection of pest population data, factors such as manual counting errors, extreme weather events, and momentary equipment malfunctions can introduce outliers that deviate from the true population dynamics, interfering with the model's prediction accuracy. The processor employs a Savitzky-Golay filter to smooth and denoise the completed population observation time series. The Savitzky-Golay filter performs low-order polynomial fitting on the data within a sliding window, replacing the original values ​​with the polynomial fit value at the window's center point. This effectively smooths the data and suppresses high-frequency random noise while preserving key information such as the signal's higher-order statistical characteristics and trend transitions. This filter is particularly suitable for biological population dynamic time series exhibiting periodic and sudden characteristics.

[0032] Furthermore, time series data often exhibit non-stationary characteristics due to temporal variations in distribution; for example, the population size of pests varies significantly across different years or seasons. This non-stationarity reduces the training efficiency and predictive performance of the model. The processor employs Reversible Instance Normalization (RevIN) to process the smoothed and denoised time series. The normalization workflow of Reversible Instance Normalization includes two reversible stages: The first stage is the positive transformation stage: the processor standardizes each time series instance separately. By subtracting the mean of the instance and dividing by the standard deviation, the sequence is transformed into a standard distribution with a mean of 0 and a variance of 1, thereby reducing the impact of non-stationary fluctuations on model learning.

[0033] The second stage is the inverse transformation stage: After the model completes the prediction and outputs the prediction sequence, the processor multiplies the prediction result by the original standard deviation and adds the original mean back to restore the numerical scale of the original feature domain, ensuring that the final prediction result is interpretable within the original feature domain.

[0034] Because the original meteorological variables are highly dimensional, and some of them have weak correlations with pest population dynamics, inputting all of them into the model would increase unnecessary computational burden and introduce redundant noise. In one embodiment of this application, the processor uses a system dual-filtering method to select covariates that are significantly correlated with the target pest population dynamics from the original multidimensional meteorological variables.

[0035] The first filter is a statistical correlation filter: the processor calculates the Spearman correlation coefficient between each meteorological variable and the population size observation time series, captures the nonlinear dependence between the target pest time series and potential covariates, and removes variables whose absolute correlation coefficient value is lower than a preset threshold.

[0036] The second filtering is a biological rationality assessment: For candidate meteorological variables that have passed the statistical correlation filter, the processor further combines the biological characteristics of the target pests for a second screening, combining the selection of meteorological variables with the biological characteristics of the pests to ensure that the selected covariates have ecological significance.

[0037] Through the above dual filtering, the processor selects meteorological covariates that are significantly related to the population dynamics of the target pest and have ecological significance from the original multidimensional meteorological variables, and generates a simplified multidimensional meteorological covariate time series as the input for the subsequent cross-correlation variable fusion step.

[0038] Step 102: Cross-correlation variables are fused between the population size observation time series and the multidimensional meteorological covariate time series to generate a fused feature sequence.

[0039] In one embodiment, the processor performs cross-correlation variable fusion on the population size observation time series and the multidimensional meteorological covariate time series to generate a fused feature sequence, including: concatenating and splicing the population size observation time series and the multidimensional meteorological covariate time series along the feature dimension to obtain a joint matrix; performing a one-dimensional convolution operation on the joint matrix along the time dimension to extract the co-evolution features between endogenous and exogenous variables and generate a cross-correlation auxiliary feature representation; and using learnable scalar parameters, with the population size observation time series as the dominant factor, weighting and summing the residuals of the population size observation time series and the cross-correlation auxiliary feature representation to output the fused feature sequence.

[0040] Using learnable scalar parameters, with the population size observation time series as the dominant sequence, the weighted residuals of the population size observation time series and the cross-correlation auxiliary feature representation are added together to output a fused feature sequence. The process includes: the processor using the population size observation time series as the dominant sequence and the cross-correlation auxiliary feature representation as auxiliary information; initializing the learnable scalar parameters, which control the information ratio between the population size observation time series and the cross-correlation auxiliary feature representation during the fusion process; calculating the product of the learnable scalar parameters and the population size observation time series to obtain the dominant weighted feature; calculating the product of the cross-correlation auxiliary feature representation and the target coefficient to obtain the auxiliary weighted feature, where the target coefficient is the difference between 1 and the learnable scalar parameters; and adding the dominant weighted feature and the auxiliary weighted feature to obtain the fused feature sequence.

[0041] Specifically, cascaded concatenation involves horizontally joining two time series along the feature dimension. Assuming the population size observation time series has a dimension of L×1, where L is the time length, and the multidimensional meteorological covariate time series has a dimension of L×d, where d is the number of meteorological covariate dimensions, then the dimension of the concatenated joint matrix is ​​L×(1+d). Each row in the joint matrix corresponds to a time point, containing the observed pest population size and all observed meteorological covariate values ​​for that time point.

[0042] In this embodiment, the one-dimensional convolution operation employs a one-dimensional convolution kernel sliding along the time dimension to capture the local co-evolution patterns of endogenous and exogenous variables on the time axis. The size of the one-dimensional convolution kernel is c×1, where c is the length of the time window, representing the number of consecutive time steps covered by a single convolution operation. For example, the kernel size c is 3 to 7, meaning that co-evolutionary features between variables are extracted within a window of 3 to 7 consecutive time steps to cover the short-term lag effect of pest population response to meteorological conditions. Endogenous variables refer to the observed time series of the target pest population itself, i.e., the population change over time, reflecting the inherent biological rhythms of the pests themselves, such as generational succession, reproduction, and death. Exogenous variables refer to the time series of multidimensional meteorological covariates, including external environmental factors such as temperature, humidity, and precipitation, reflecting the external conditions affecting the dynamics of pest populations.

[0043] The cross-correlation auxiliary feature representation is a feature sequence extracted from the joint matrix through one-dimensional convolution. The dimension of this feature sequence is the same as that of the population size observation time series, both being L×1. That is, each time point corresponds to a comprehensive feature value that integrates meteorological information and population information within the preceding and following time windows. The core function of this representation is to capture the cross-coupling information of how meteorological changes affect pest populations.

[0044] The learnable scalar parameter, denoted as α, is a scalar value that is automatically optimized and updated during model training using gradient descent. Its value ranges from 0 to 1. The initial value of α can be set to 0.9, indicating that the model initially relies primarily on information from the pest population observation time series. In the initial training phase, when α = 0.9, 90% of the information in the fused feature sequence comes from the population size observation time series. The model primarily learns based on the pest population's own evolutionary patterns, introducing only a small amount of auxiliary information from meteorological covariates (accounting for 10%). As training iterates, α adaptively adjusts within the range of 0 to 1, and the model gradually optimizes the optimal ratio between the pest's endogenous sequence and meteorological exogenous information. When training converges, α stabilizes at its optimal value, at which point the model achieves the optimal dynamic balance between the pest's endogenous sequence and meteorological exogenous information. The final value of α is automatically determined by the training data, requiring no manual setting, thus exhibiting data adaptability.

[0045] In this embodiment, the mathematical form of weighted residual summation is as follows: ,in, Represents the fused feature sequence. This represents the time series of population size observations. This represents a multidimensional meteorological covariate time series, and Concat(·) represents a cascade concatenation operation. (·) denotes a one-dimensional convolution extraction function. The parameter is a learnable scalar parameter. In the weighted residual summation architecture, the weight of the population size observation time series is α, and the weight of the cross-correlation auxiliary feature representation is (1-α). Through adaptive adjustment of α within the range of 0 to 1, the model dynamically balances the ratio of its own evolutionary information and meteorological auxiliary information. When the meteorological information in the data has a strong explanatory power for the pest population, α automatically tends to a smaller value, and (1-α) tends to a larger value, increasing the proportion of meteorological information injection; when the meteorological information is redundant or noisy, α automatically tends to a larger value, and (1-α) tends to a smaller value, suppressing the injection of meteorological information. In the embodiment of this application, the initial value of α is set to 0.9, so that the model can fully learn the temporal autocorrelation of the pest population itself in the early stage of training, and then gradually optimize the fusion ratio as the training process progresses, thereby obtaining a more stable training process and better final prediction performance.

[0046] Through the above-described cross-correlation variable fusion steps, a fused feature sequence is output. Its dimension is L×1, consistent with the dimension of the original population size observation time series. This fused feature sequence retains the core evolutionary information of the pest population while incorporating effective information from meteorological covariates in an adaptive weighted manner, providing high-quality univariate feature input for subsequent periodic extraction and block prediction.

[0047] Step 103: Perform frequency domain transformation on the fused feature sequence, extract the core physical period corresponding to at least one frequency component whose energy intensity meets the preset conditions, and form a period set.

[0048] The processor performs a frequency domain transformation on the fused feature sequence, extracting the core physical cycles corresponding to at least one frequency component whose energy intensity meets preset conditions, forming a cycle set. This includes: performing a fast Fourier transform on the fused feature sequence to calculate the amplitude value of each frequency component; setting an effective cycle interval based on the biological reproduction characteristics of the target pest, and filtering out frequency components outside the effective cycle interval; sorting the frequency components retained within the effective cycle interval according to their amplitude values ​​from largest to smallest; determining the top K frequency components after sorting as selected frequency components, where K is a preset positive integer; and extracting the cycle corresponding to each selected frequency component to form a cycle set.

[0049] In this embodiment, the frequency domain transformation is the Fast Fourier Transform (FFT). FFT is an efficient algorithm implementation of the Discrete Fourier Transform, suitable for time series processing scenarios with a relatively large length L (e.g., 180 to 730 days) as described in this application, ensuring the real-time performance and computational efficiency of the period extraction step. Optionally, the frequency domain transformation can also be implemented using methods such as Discrete Fourier Transform, wavelet transform, or autocorrelation analysis.

[0050] Specifically, the Fast Fourier Transform will fuse feature sequences. The time-domain to frequency-domain transformation decomposes a discrete-time signal of length L into a superposition of L frequency components. Each frequency component corresponds to a period and an amplitude value, where the amplitude value reflects the energy intensity of that frequency component throughout the time series. A larger amplitude value indicates that the frequency component plays a more dominant role in the periodic fluctuations of the sequence. In practical applications, since the fused feature sequence is a real-number sequence, the spectrum obtained after the Fast Fourier Transform (FFT) exhibits conjugate symmetry. Therefore, only the first half of the frequency domain needs to be analyzed to obtain information on all effective frequency components.

[0051] The mathematical expression for the Fast Fourier Transform is: Where F(k) is the complex representation of the k-th frequency component, and k is the frequency index. Let j be the time index, j be the imaginary unit, and L be the sequence length. The amplitude of each frequency component is A(k) = |F(k)|, which is the modulus represented by the complex number.

[0052] The effective periodicity range is set based on the biological generation cycle of the target pest. Taking the rice stem borer as an example, it typically has 2 to 3 generations per year, with a complete generation cycle taking approximately 30 to 60 days. Annual population fluctuations are influenced by overlapping generations, exhibiting multiple periodic components. Therefore, for the rice stem borer, the lower limit of the effective periodicity range is 3 to 5 days to exclude high-frequency pseudo-cycles caused by sensor noise or random fluctuations; the upper limit is 100 to 120 days to exclude extremely low-frequency components that exceed the reasonable biological range of the pest. This range can be adjusted for different pest species based on their biological characteristics. For example, for pests with shorter generation cycles, such as aphids, the upper limit of the effective periodicity range can be set to 60 to 90 days.

[0053] Sort by amplitude value from largest to smallest. This means arranging all frequency components within the effective period range in descending order of their amplitude values, so that the frequency component with the strongest energy is placed first, and so on.

[0054] The preset positive integer K represents the preset number of frequency components to select, serving as a manually set hyperparameter for the model. Specifically, the processor starts with the frequency component with the largest amplitude and selects the first K frequency components in descending order of amplitude value, defining these K frequency components as the selected frequency components. This selection rule ensures that the selected K frequency components have the strongest energy among all effective frequency components, meaning they dominate the periodic fluctuations of the sequence.

[0055] For example, suppose 20 frequency components are retained after filtering the effective period intervals. The processor sorts them from largest to smallest amplitude: the first component has an amplitude of 40, the second 25, the third 15, the fourth 10, the fifth 5, and the sixth and subsequent components all have amplitudes less than 5. If the preset positive integer K=4, the processor selects the first four frequency components with the largest amplitudes as the selected frequency components; if the preset positive integer K=5, the processor selects the first five frequency components with the largest amplitudes as the selected frequency components. In this way, the value of K is preset manually according to the actual prediction task requirements, and the selected frequency components are always the K most energetic biological rhythm cycles.

[0056] Then, the processor extracts the period values ​​corresponding to each selected frequency component, forming a period set. The correspondence between frequency components and periods is: period P = L / k, where L is the sequence length and k is the frequency index. For each of the selected m frequency components, its corresponding period values ​​form a period set { }, where m is a natural number, automatically determined by the cumulative contribution rate screening rule mentioned above. These cycles are the core physical cycles, and their physical meaning is the length of the m biological rhythm cycles in the fused feature sequence where energy is most concentrated.

[0057] The core physical cycle is the most energy-concentrated biological rhythm cycle in the fused feature sequence. For agricultural pests such as the rice stem borer, the core physical cycle typically corresponds to multiple fluctuations in pest population size throughout the year with seasonal changes, possibly including: an annual cycle of approximately 365 days, a semi-annual cycle of approximately 180 days, and a generation cycle of approximately 30 to 60 days. These cycles directly reflect the intrinsic rhythms of pest population dynamics, providing a data-driven, biologically meaningful cyclical basis for subsequent scale-aligned segmentation.

[0058] Step 104: Based on the period set, dynamically resample the fused feature sequence to generate multiple sets of adaptive block matrices; wherein, each set of adaptive block matrices corresponds to a core physical period in the period set, and the block scale in each set of adaptive block matrices is associated with the core physical period corresponding to that set.

[0059] The processor dynamically resamples the fused feature sequence based on the period set to generate multiple sets of adaptive block matrices. This includes: calculating the ideal stretch length for each core physical cycle in the period set; using a linear interpolation algorithm to resample the fused feature sequence to the ideal stretch length to generate an interpolation sequence; dividing the interpolation sequence into a preset number of non-overlapping blocks, with the physical time sequence length of each block equal to the currently processed core physical cycle; and reshaping the dimensions of all the equally divided blocks to generate a set of adaptive block matrices corresponding to the core physical cycle.

[0060] Specifically, for each core physical cycle in the cycle set, its ideal stretch length is calculated, including: obtaining the preset total number of blocks, the cycle duration of the current core physical cycle, and the original timing length of the fused feature sequence; the processor uses the product of the preset total number of blocks and the cycle duration of the current core physical cycle as the theoretical sequence length; the processor compares the numerical values ​​of the theoretical sequence length and the original timing length of the fused feature sequence, and selects the larger of the two as the ideal stretch length corresponding to the current core physical cycle.

[0061] More specifically, for periodic sets { The i-th core physics cycle in} The formula for calculating its ideal tensile length is: Where N is the preset total number of blocks, representing the number of non-overlapping blocks obtained by uniformly dividing the interpolation sequence; The current core physical cycle is represented by the number of time steps; L is the original time series length of the fused feature sequence. When the value is greater than L, the ideal stretch length is At this point, the interpolated sequence is stretched (length increased) relative to the original sequence, so that the physical duration of each block is exactly equal to... When N× When the value is less than or equal to L, the ideal stretching length is L. At this time, the length of the interpolated sequence is equal to the length of the original sequence or slightly compressed, ensuring that even if the core physical period is short, the interpolated sequence will not be over-compressed and cause information loss. The max operation guarantees the stability and information integrity of the sequence during dynamic resampling.

[0062] The preset total number of blocks N is a pre-defined positive integer, representing the number of blocks obtained by dividing the interpolation sequence for each set of adaptive block matrices. In this embodiment, the value of N corresponds to the length of the attention vector in the subsequent block-level attention weighting module. The value of N ranges from 8 to 32, and its value determines the granularity of the blocks: the larger the value of N, the fewer time steps each block contains, the finer the block granularity, and the more precisely it can capture local temporal features; the smaller the value of N, the more time steps each block contains, the coarser the block granularity, and the better it can preserve global context information.

[0063] In this embodiment, the processor uses a linear interpolation algorithm to resample the fused feature sequence to an ideal stretched length. Specifically, it is assumed that the original fused feature sequence... The length is L, and the target ideal stretch length is... The processor constructs a length of The target time axis is used to map interpolation points on the target time axis to their corresponding positions on the original time axis. The linear relationship between two adjacent known data points in the original sequence is used to estimate the value of the interpolation point. For the q-th interpolation point on the target time axis, its corresponding position on the original time axis is pos = q×(L-1) / ( -1), select two adjacent integer positions before and after pos. = floor(pos) and If = ceil(pos), then the value of the interpolation point is: ,in, For interpolation sequences, Here, is the original fused feature sequence, pos is the mapping position of the interpolation point in the original sequence, and floor and ceil are the floor function and floor function, respectively. Through the above linear interpolation method, the temporal length of the interpolated sequence is precisely adjusted to the ideal stretch length. Meanwhile, the numerical trend in the sequence is smoothly maintained, avoiding spurious fluctuations that may be introduced by nonlinear interpolation.

[0064] Furthermore, the processor will have a length of The interpolated sequence is divided into N consecutive non-overlapping segments along the time dimension, each segment being called a block. Due to the ideal stretch length... Through N× The result is obtained by taking the maximum value of L, when N× When the value is greater than L, each block contains exactly L. One time step; when N× When L is less than or equal to L, each block contains L / N time steps, and its corresponding physical time length is L / N, which is not less than L / N. This ensures that even during the cycle A smaller size will not lead to over-compression. The physical timing length of each block is equal to the current core physical cycle. That is, the time span corresponding to each block on the original timeline is exactly 1 / 3. Each time unit.

[0065] For example, suppose the length of the original fused feature sequence is L=180 days, the preset total number of blocks is N=12, and the period set contains one core physical period. =30 days. Therefore, the theoretical sequence length is N× =12 × 30 = 360 days, which is greater than the original length of 180 days, representing the ideal stretching length. =max(360, 180) = 360 days. The processor stretches the fused feature sequence from 180 days to 360 days using linear interpolation, then divides it into 12 blocks, each containing 30 time steps (30 days), corresponding exactly to a complete 30-day generation cycle. If another core physical cycle... =60 days, then the ideal elongation length =max(12×60,180)=max(720,180)=720 days, each block contains 60 time steps (60 days), corresponding to a 60-day generation cycle.

[0066] In this embodiment, the processor stacks N blocks in chronological order to construct an N×S matrix, where N is the preset total number of blocks, and S is the number of time steps contained in each block after linear interpolation stretching. =N× At that time, S= That is, the matrix dimension is N× ;when When S = L, S = L / N, meaning the matrix dimension is N×(L / N). Each block is a row of this matrix, and the element in the j-th row and s-th column of the matrix corresponds to the fused feature value at the s-th time step in the j-th block.

[0067] Through the above processing, for each core physical cycle in the cycle set Each generates a corresponding set of adaptive block matrices. ,in The number of time steps for each block in the i-th group of matrices (equal to the ideal stretch length corresponding to that group divided by the preset total number of blocks). The entire periodic set generates m groups of adaptive block matrices. }

[0068] It is important to emphasize that the adaptability of this step is reflected in two aspects: First, the total number of blocks N is fixed, but because the ideal stretching length corresponding to each group of matrices is different, the actual physical scale of each block in each group of matrices, i.e., the number of time steps contained in each block, is related to the corresponding core physical cycle. The block scales of different groups are different, thus achieving scale adaptation. Second, dynamic resampling ensures that even within the same group of matrices, the block scale is dynamically adjusted according to the corresponding core physical cycle, rather than using a fixed block length. This dynamic resampling mechanism ensures that each group of adaptive block matrices has the same number of rows, facilitating subsequent unified processing; and the number of columns in each row, i.e., the actual length of each block, is related to the corresponding core physical cycle, ensuring that the physical time span of each block fully covers a core physical cycle, thereby preserving the natural semantic integrity of the pest population reproduction cycle and avoiding cycle fragmentation caused by fixed blocks.

[0069] Through step 104 above, the fused feature sequence is transformed into m sets of adaptive block matrices. Each set of matrices corresponds to a core physical cycle. The scale of each block in each set of matrices is associated with the corresponding core physical cycle, providing multi-scale, physically meaningful block feature inputs for subsequent dual-weighted fusion.

[0070] In one embodiment of this application, in addition to linear interpolation, resampling may also employ any one of cubic spline interpolation, piecewise cubic Hermite interpolation, or nearest neighbor interpolation. Linear interpolation is preferred because it is computationally efficient and can effectively avoid interpolation artifacts such as overshoot and oscillation.

[0071] In one embodiment of this application, after generating the adaptive block matrix, the processor also generates a corresponding physical time identifier for each block and records the start and end positions of each block in the original fused feature sequence, so that the attention weights can be correlated with specific time periods during subsequent block-level attention weighting, thereby accurately identifying the key precursor window before the outbreak of pests.

[0072] Step 105: Perform double-weighted fusion on multiple sets of adaptive block matrices to output a comprehensive feature tensor.

[0073] The processor performs dual-weighted fusion on multiple sets of adaptive block matrices to output a comprehensive feature tensor. This includes: performing adaptive average pooling on the interpolation sequence corresponding to the current scale, compressing the temporal length of the interpolation sequence back to the original length of the fused feature sequence to obtain a pooled sequence; flattening the pooled sequence into a one-dimensional vector and inputting it into a linear evaluation network to output an attention weight vector with a dimension equal to the preset total number of blocks, and then normalizing the attention weight vector using a normalized exponential function; multiplying each normalized weight scalar into each temporal block to complete the time-dimensional differentiated feature focusing; performing adaptive average pooling again on all weighted temporal blocks to uniformly compress all blocks to a fixed standard block length and outputting a standardized feature matrix; and performing equal-weighted superposition fusion on the standardized feature matrices obtained at different core physical cycle scales to obtain a comprehensive feature tensor.

[0074] Specifically, dual-weighted fusion refers to performing two different levels of weighted fusion operations on multiple sets of adaptive block matrices: the first level is patch-level attention weighting, which assigns differentiated weights to each block within each set of adaptive block matrices, focusing on the key precursor window before pest outbreaks; the second level is multi-scale equal-weight integration, which performs equal-weighted fusion between the standardized feature matrices corresponding to different core physical cycles, combining the evolutionary advantages of long and short cycles.

[0075] The dual-weighted fusion employs a processing logic of first assigning differential weights at the block level, and then performing equal-weighted integration across scales. Block-level attention weighting is responsible for highlighting key time periods in the time dimension. Its mechanism applies between the N blocks of each set of adaptive block matrices, enabling differentiated processing of information from different time periods within a cycle. Multi-scale equal-weighted integration is responsible for integrating information across different cycle scales. Its mechanism applies between m sets of standardized feature matrices, enabling the equal-weighted fusion of features from long and short cycles.

[0076] The processor performs adaptive average pooling on the interpolation sequence corresponding to the current scale, compressing the temporal length of the interpolation sequence back to the original length of the fused feature sequence to obtain the pooled sequence. Specifically, for the i-th core physical cycle in the cycle set... The processor obtains the interpolation sequence generated in step 104, with a length of The interpolation sequence is obtained by stretching or compressing the original fused feature sequence, and its time length is... Unlike the original length L, to evaluate the importance weights of each block, the interpolated sequence needs to be mapped back to the original time axis. The processor performs adaptive average pooling on this interpolated sequence, reducing the length to L. The interpolation sequence is compressed to a length of L, which is the original length of the fused feature sequence, to obtain the pooling sequence. The temporal resolution of this pooling sequence is consistent with the original length L of the fused feature sequence, so that the attention weights calculated subsequently can accurately correspond to specific time periods on the original time axis.

[0077] For example, the original fusion feature sequence length is L=180 days, and a certain core physical cycle The corresponding interpolation sequence length is =360 days. Adaptive average pooling compresses the 360-day interpolated sequence to 180 days, ensuring that each time point in the pooled sequence corresponds to the information of a time point in the original sequence. In one embodiment, adaptive average pooling employs a one-dimensional adaptive average pooling operation. By setting the output length to L, the input sequence is divided into L intervals, and the average value within each interval is taken as the pooling output value for that interval. Since L is the length of the original sequence, the pooled sequence maintains the same temporal resolution as the original sequence.

[0078] The processor flattens the pooling sequence into a one-dimensional vector and inputs it into the linear evaluation network. The output is an attention weight vector with a dimension equal to the preset total number of blocks. This attention weight vector is then normalized using a normalized exponential function. Specifically, the processor flattens a pooling sequence of length L into a one-dimensional vector and inputs this vector into the linear evaluation network. The linear evaluation network is a neural network module containing a linear layer (fully connected layer), with an input dimension of L and an output dimension of N, where N is the preset total number of blocks. The learnable weight parameters of the linear evaluation network map the L-dimensional pooling sequence to an N-dimensional original attention vector, with each dimension corresponding to the importance evaluation score of a block.

[0079] To transform the original attention vector into weights in a usable probability distribution form, the processor performs a normalized exponential function (Softmax) operation on the original attention vector, so that all N weight values ​​are positive and their sum is 1.

[0080] The normalized weight scalars are multiplied one-to-one into each time series block to achieve time-dimensional differential feature focusing. Specifically, the processor uses the N time series blocks generated in step 104, each block having a length of... Each of these is then multiplied element-wise by the normalized N attention weight scalars. Specifically, all time steps in the j-th block are multiplied by the weights. That is, the feature value of each time step in the j-th block is multiplied by the same weight. Mathematically, this is represented as: ,in, For the j-th original block, This is the j-th block after weighting.

[0081] Through the above operations, high-importance segments, such as the pre-outbreak period of a sharp increase in pest numbers, are assigned greater weight, and their features are amplified in subsequent processing; low-importance segments, such as the overwintering period when pest numbers remain stable for a long time, are assigned less weight, and their features are suppressed. This mechanism achieves differentiated feature focusing in the time dimension, enabling the model to accurately identify and enhance the feature expression of key pre-outbreak windows.

[0082] In this embodiment, adaptive average pooling is performed again on all time-series blocks after weighted processing to compress all blocks to a fixed standard block length, outputting a standardized feature matrix. Specifically, the processor performs adaptive average pooling on each of the N weighted blocks, compressing all blocks to the same fixed standard block length C. The fixed standard block length C is a pre-set positive integer representing the number of time steps contained in each standardized block. Since the block lengths corresponding to different core physical cycles in step 104... After being compressed to the same standard block length C, the feature matrices at different periodic scales have the same dimension, which facilitates subsequent equal-weight fusion operations.

[0083] Adaptive average pooling will divide each length into several equal parts. The block sequence is compressed to a length of C. For the j-th weighted block in the i-th adaptive block matrix, pooling yields a normalized block of length C. The N normalized blocks are stacked in chronological order to form a normalized feature matrix of dimension N×C. This normalized feature matrix has the same number of rows as the original block matrix (both are N), but the number of columns in each row is uniformly C, achieving dimensional uniformity of the feature matrix at different periodic scales.

[0084] Through the aforementioned first-level block-level attention weighting, for each of the m core physical cycles... Each generates a corresponding set of standardized feature matrices.

[0085] Furthermore, the processor performs equal-weighted superposition and fusion on the standardized feature matrices obtained at different core physical cycle scales to obtain a comprehensive feature tensor. Specifically, the processor performs equal-weighted averaging fusion on m sets of standardized feature matrices along the feature dimension. Since all standardized feature matrices have an N×C dimension, element-wise averaging can be performed directly along the feature dimension. Multi-scale equal-weighted ensemble draws on the idea of ​​ensemble learning to perform equal-weighted ensemble on features extracted at different core physical cycle scales. Since different core physical cycles characterize the evolutionary patterns of pest populations at different time scales, equal-weighted ensemble abandons the absolute bias towards a single dominant cycle and instead equally integrates feature information from each scale: short-cycle features can accurately capture the rapid fluctuations of pest populations within generations, while long-cycle features can robustly characterize the macro-trends of the population throughout the year. Equal-weighted ensemble complements the advantages of both, improving the model's generalization ability and robustness under complex weather conditions.

[0086] In one embodiment, the mathematical expression for equal-weighted averaging is: Where m is the total number of core physical cycles in the period set, i.e., the number of selected frequency components. Let i be the normalized feature matrix corresponding to the i-th core physical cycle. This is the comprehensive feature tensor output after equal weight averaging.

[0087] Through the aforementioned second-level multi-scale weighted integration, the processor outputs a comprehensive feature tensor with dimensions N×C. This integrated feature tensor undergoes both block-level attention-weighted differential feature focusing and multi-scale equal-weighted information integration, providing high-quality feature input for subsequent prediction output.

[0088] In one embodiment of this application, Softmax normalization can be achieved using Softmax with a temperature coefficient, specifically expressed as follows: , This is the temperature coefficient. When... When the value is greater than 1, the Softmax output is smoother, and the difference in weights between blocks decreases; when... When the value is less than 1, the Softmax output is sharper, and the weight differences between each block increase. The concentration of attention weights can be controlled by adjusting the temperature coefficient. Preferably, =1, which is the standard Softmax.

[0089] In one embodiment, multi-scale equal-weighted ensemble can also use weighted summation instead of equal-weighted averaging, that is, assigning different fusion weights to the standardized feature matrices corresponding to different core physical cycles, and the fusion weights are automatically learned during model training. However, using equal-weighted ensemble can avoid the model's over-reliance on a single cycle scale, enhance generalization ability, and achieve effective integration of multi-scale information without the need for additional learnable parameters.

[0090] Step 106: Based on the comprehensive feature tensor, make predictions and output the future pest population size prediction sequence.

[0091] The processor makes predictions based on the comprehensive feature tensor and outputs a future pest population size prediction sequence. This process also includes: flattening the comprehensive feature tensor along all dimensions to obtain a one-dimensional global context vector; inputting the one-dimensional global context vector into a multi-layer fully connected network, which contains multiple sequentially connected fully connected layers, with adjacent fully connected layers undergoing non-linear transformation via Gaussian error linear units; directly mapping the one-dimensional global context vector through the multi-layer fully connected network to convert it into an output vector with a dimension equal to the target prediction horizon step size; and using the output vector as the future pest population size prediction sequence within a preset time period.

[0092] Specifically, the comprehensive feature tensor is flattened along all dimensions to obtain a one-dimensional global context vector, including: the processor acquiring the comprehensive feature tensor output in step 105. The tensor has dimensions N×C, where N is the total number of preset blocks and C is the fixed standard block length. The processor performs a flattening operation along all dimensions of the tensor, that is, unfolds all elements of the N×C two-dimensional tensor sequentially to form a one-dimensional vector of length N×C. This one-dimensional vector integrates the feature information of all N blocks at each time step, and contains the complete feature expression of the pest population after differential weighting at different core physical cycle scales, hence it is called a one-dimensional global context vector.

[0093] In the N×C dimension of the comprehensive feature tensor, N is usually between 8 and 32, and C is usually between 4 and 32. Therefore, the length of the one-dimensional global context vector is usually between 32 and 1024. This vector serves as the feature of the input layer of the subsequent multi-layer fully connected network, and its dimension is the input dimension of the multi-layer fully connected network.

[0094] Furthermore, the processor takes the one-dimensional global context vector as input to a multi-layer fully connected network, which is then processed through a network structure consisting of multiple fully connected layers connected sequentially. Each fully connected layer performs a linear transformation, multiplying the input feature vector by a learnable weight matrix and adding a bias term. A Gaussian Error Linear Unit (GELU) activation function is placed between adjacent fully connected layers to introduce a non-linear transformation.

[0095] The hidden layer dimensions of each fully connected layer decrease layer by layer to progressively compress the feature dimensions and extract high-level semantic information. This progressively decreasing dimension allows the network to gradually extract the most essential predictive information from high-dimensional features, while reducing the number of model parameters and mitigating the risk of overfitting.

[0096] Adjacent fully connected layers undergo nonlinear transformation using the GELU activation function, while the last layer, i.e. the output layer, does not have an activation function set and directly outputs a vector with the same dimension as the target prediction horizon step size, so as to ensure that the predicted value can freely take values ​​in the real number range.

[0097] The output layer dimension of a multilayer fully connected network is set to , The step size for predicting the target horizon represents the number of future time steps to be predicted. For example, if we need to predict the daily pest population size for the next 7 days, then... =7; if you need to predict the next 14 days, then =14. Through layer-by-layer mapping of a multi-layer fully connected network, the input one-dimensional global context vector (dimension N×C) is transformed into a vector of length 14. The output vector.

[0098] The processor will output the h-th element (h=1,2,...) in the vector. As the predicted pest population size at the h-th time step, all These elements constitute a complete sequence for predicting the future population size of pests.

[0099] It should be noted that the prediction strategy used in this step is Direct Multi-step Forecasting, which means that all predictions are output at once through a multi-layer fully connected network. This approach outputs predictions for all future time steps in a single forward propagation. This contrasts sharply with traditional single-step rolling prediction strategies, where the model predicts only one future time step at a time, then uses that prediction as part of the input to iteratively predict the next time step, repeating this process until the target prediction horizon is reached. This method suffers from significant error accumulation; the prediction error of the previous step is amplified with each iteration, ultimately leading to a sharp decline in long-term prediction accuracy. In contrast, the direct multi-step prediction strategy employed in this application outputs predictions for all future time steps in a single forward propagation. Predictions for each time step are generated simultaneously and independently, effectively avoiding error accumulation. This approach is particularly suitable for time series prediction tasks involving pest populations, which exhibit complex nonlinear dynamic characteristics.

[0100] In the above technical solution, this application integrates cross-correlated variables, with the pest population sequence as the main component to integrate meteorological covariate information, effectively suppressing the interference of redundant meteorological noise on the main evolutionary line of the pests themselves, so that the model focuses on the essential laws of the target variable; secondly, it adaptively extracts the core physical cycle with the strongest energy through frequency domain transformation, and dynamically resamples the fused feature sequence according to the cycle to generate an adaptive block matrix, so that the block scale is related to the real biological rhythm cycle of the pest population, avoiding the fragmentation of the biological evolution cycle caused by fixed blocks, and preserving the natural continuity and semantic integrity of the temporal information; finally, through dual weighted fusion, it uses a block-level attention mechanism to differentiate the weights of each block in the time dimension, accurately locks the key precursor window before the pest outbreak, and integrates the evolutionary features under different physical cycle scales, taking into account the advantages of long and short cycles, and improving the generalization ability and robustness of the model in complex meteorological environments.

[0101] This application provides a Fourier adaptive variable fusion time series prediction device, the device comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the Fourier adaptive variable fusion time series prediction method as described above.

[0102] This application provides a storage medium storing a program that, when executed by a processor, implements any of the above-described Fourier adaptive variable fusion time series prediction methods.

[0103] This application provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the above-described Fourier adaptive variable fusion time series prediction methods. The device described herein can be a server, PC, PAD, mobile phone, etc.

[0104] This application also provides a computer program product that, when executed on a data processing device, is adapted to execute a program that initializes a Fourier adaptive variable fusion time series forecasting method having any of the steps described above.

[0105] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0110] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0111] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0112] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0113] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A variable fusion time series forecasting method based on Fourier adaptive variability fusion, characterized in that, The method includes: Obtain the observation time series of the population size of the target pest within a historical period, and obtain the time series of multidimensional meteorological covariates concurrent with the historical period. The time series of population size observations and the time series of multidimensional meteorological covariates are cross-correlated variables fused to generate a fused feature sequence. Perform frequency domain transformation on the fused feature sequence to extract the core physical period corresponding to at least one frequency component whose energy intensity satisfies the preset condition, and form a period set; Based on the set of periods, the fused feature sequence is dynamically resampled to generate multiple sets of adaptive block matrices; wherein, each set of adaptive block matrices corresponds to a core physical period in the set of periods, and the block scale in each set of adaptive block matrices is associated with the core physical period corresponding to that set. Perform double-weighted fusion on the multiple sets of adaptive block matrices to output a comprehensive feature tensor; Based on the comprehensive feature tensor, a prediction sequence for the future pest population size is output.

2. The method according to claim 1, characterized in that, The process involves performing a frequency domain transformation on the fused feature sequence to extract the core physical periods corresponding to at least one frequency component whose energy intensity satisfies a preset condition, thereby forming a period set, including: Perform a Fast Fourier Transform on the fused feature sequence to calculate the amplitude value of each frequency component; The effective period range is set according to the biological reproduction characteristics of the target pest, and frequency components outside the effective period range are filtered out. The frequency components retained within the effective period interval are sorted from largest to smallest according to their amplitude values; The first K frequency components after sorting are determined as the selected frequency components, where K is a preset positive integer; The periods corresponding to each selected frequency component are extracted to form the period set.

3. The method according to claim 1, characterized in that, The step of dynamically resampling the fused feature sequence according to the period set to generate multiple sets of adaptive block matrices includes: For each core physical cycle in the aforementioned cycle set, calculate its ideal stretch length; A linear interpolation algorithm is used to resample the fused feature sequence to the ideal stretching length to generate an interpolated sequence; The interpolation sequence is divided into a preset total number of non-overlapping blocks, and the physical time length of each block is equal to the core physical cycle of the current processing. The dimensions of all the evenly divided blocks are reshaped to generate a set of adaptive block matrices corresponding to the core physical cycle.

4. The method according to claim 3, characterized in that, The calculation of the ideal stretch length for each core physical cycle in the set of cycles includes: Obtain the preset total number of blocks, the cycle duration of the current core physical cycle, and the original time sequence length of the fused feature sequence; The product of the preset total number of blocks and the cycle duration of the current core physical cycle is used as the theoretical sequence length; By comparing the theoretical sequence length with the original time-series length of the fused feature sequence, the larger of the two values ​​is selected as the ideal stretch length corresponding to the current core physical cycle.

5. The method according to claim 1, characterized in that, The step of predicting the future pest population size based on the comprehensive feature tensor and outputting a predicted sequence includes: Flatten the comprehensive feature tensor along all dimensions to obtain a one-dimensional global context vector; The one-dimensional global context vector is input into a multi-layer fully connected network, which contains multiple fully connected layers connected in sequence, and adjacent fully connected layers undergo nonlinear transformation through Gaussian error linear units. The one-dimensional global context vector is directly mapped through the multi-layer fully connected network and converted into an output vector with the same dimension as the target prediction horizon step size. The output vector is used as a prediction sequence for the pest population size within a preset time period in the future.

6. The method according to claim 1, characterized in that, The process of cross-correlation variable fusion of the population size observation time series and the multidimensional meteorological covariate time series to generate a fused feature sequence includes: Along the feature dimension, the population size observation time series and the multidimensional meteorological covariate time series are concatenated and spliced ​​to obtain a joint matrix; A one-dimensional convolution operation is performed on the joint matrix along the time dimension to extract the co-evolution features between endogenous and exogenous variables and generate cross-correlation auxiliary feature representations. Using learnable scalar parameters, with the population size observation time series as the main component, the population size observation time series and the cross-correlation auxiliary feature representation are weighted and summed to output the fused feature sequence.

7. The method according to claim 6, characterized in that, The method utilizes learnable scalar parameters, with the population size observation time series as the primary factor, and performs weighted residual addition of the population size observation time series and the cross-correlation auxiliary feature representation to output a fused feature sequence, including: The population size observation time series is used as the dominant sequence, and the cross-correlation auxiliary feature representation is used as auxiliary information; Initialize learnable scalar parameters, which are used to control the information ratio between the population size observation time series and the cross-correlation auxiliary feature representation during the fusion process; The dominant weighted feature is obtained by multiplying the learnable scalar parameter with the population size observation time series. The product of the cross-correlation auxiliary feature representation and the target coefficient is calculated to obtain the auxiliary weighted feature, wherein the target coefficient is the difference between 1 and the learnable scalar parameter; The dominant weighted feature is added to the auxiliary weighted feature to obtain the fused feature sequence.

8. The method according to claim 1, characterized in that, The process of performing dual-weighted fusion on the multiple sets of adaptive block matrices to output a comprehensive feature tensor includes: Adaptive average pooling is performed on the interpolation sequence corresponding to the current scale to compress the temporal length of the interpolation sequence back to the original length of the fused feature sequence, thus obtaining the pooling sequence. The pooling sequence is flattened into a one-dimensional vector and input into a linear evaluation network. The output is an attention weight vector with a dimension equal to the preset total number of blocks. The attention weight vector is then normalized using a normalized exponential function. Each normalized weight scalar is multiplied into each time series block to achieve the focus of time-dimension differentiated features. Adaptive average pooling is performed again on all time series blocks after weighted processing to compress all blocks to a fixed standard block length and output a standardized feature matrix. The standardized feature matrices obtained under different core physical cycle scales are subjected to equal-weight superposition and fusion to obtain the comprehensive feature tensor.

9. A variable fusion time series forecasting device based on Fourier adaptive variability fusion, characterized in that, The device includes: Memory, used to store computer programs; A processor, configured to implement the Fourier adaptive variable fusion time series prediction method as described in any one of claims 1 to 8 when executing the computer program.

10. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the Fourier adaptive variable fusion time series prediction method according to any one of claims 1 to 8.

Images