Agricultural insurance compensation amount prediction method based on time sequence external regression

Abstract

The invention discloses an agricultural insurance compensation amount prediction method based on time sequence external regression, and belongs to the field of agricultural insurance compensation, and the method specifically comprises the following steps: collecting meteorological index data, crop planting area data and crop loss compensation data, and constructing original data; processing the original data to generate agrometeorological time series data; constructing a time series data conversion unit, and carrying out standardization processing, filling, cutting and stacking on the agrometeorological time series data; an additional data coding unit is constructed, and external information is fused into agricultural meteorological time series data in a one-hot coding mode; and the backbone network is used for adjusting the input and output of the backbone network and predicting the agricultural insurance compensation amount. According to the method, the claim condition caused by agricultural loss possibly generated in the planting area is accurately predicted from the historical time sequence meteorological data, and an agricultural insurance company is helped to accelerate the claim process and improve the claim accuracy.

Description

Technical Field

[0001] The invention belongs to the field of agricultural insurance compensation, and specifically relates to a method for predicting agricultural insurance compensation amount based on time series external regression. Background Art

[0002] Agricultural insurance is used to mitigate the economic losses suffered by agricultural producers in weather disasters and plays a vital role in ensuring the stability of agricultural production and food security. Due to the complexity of meteorological changes and crop growth mechanisms, it is still extremely challenging to accurately predict the amount of compensation for agricultural insurance claims caused by weather events.

[0003] There are currently three main schemes for predicting the amount of agricultural insurance compensation, which are suitable for different insurance types and scenarios. The first scheme is the forecasting scheme of the amount of compensation for meteorological index insurance. This scheme relies on meteorological experts to predict claims based on meteorological models, and closely links the insurance compensation standards with specific meteorological indicators. If the meteorological indicators (such as precipitation or temperature) in a certain area reach or exceed the set threshold, the compensation conditions are met, and the compensation amount will be calculated according to the change ratio of the indicator. Since this scheme does not need to evaluate the actual crop losses, it is convenient to apply existing meteorological models and reduce the complexity and cost of data collection. Meteorological index insurance clarifies the compensation conditions at the beginning of its setting, simplifies the subsequent claims process, and is suitable for areas with frequent extreme weather. However, the limitation of this method is that it can only be applied to insurance types directly related to meteorological factors, such as index insurance for meteorological disasters such as droughts and floods, and cannot be applied to insurance based on actual crop yields and economic losses, such as cost insurance and income insurance.

[0004] The second option is to combine post-disaster yield measurement with crop model deduction and claim prediction. This option is to determine the actual damage to crops through regional yield measurement after the meteorological disaster occurs, and to combine crop growth and development models and climate influencing factors to deduce claims. Specifically, after evaluating the growth of crops after the disaster, the model will infer the possible agricultural losses in the future harvest stage based on the crop growth cycle and climate influencing factors (such as temperature, precipitation, sunshine duration, etc.), and finally calculate the insurance compensation amount. Because this method is based on actual disaster data, it can more accurately reflect the actual loss of crops after the disaster. It is currently the most mainstream method for predicting the amount of agricultural insurance compensation, and is particularly suitable for agricultural insurance scenarios in multiple categories and regions. However, due to its reliance on manual yield measurement, this method is time-consuming and labor-intensive, and the accuracy of the prediction is greatly affected by the post-disaster yield measurement data. In addition, the effect of the yield measurement model is different for different regions and crops, which may affect the prediction accuracy under complex meteorological disasters.

[0005] The third solution is to use the convolutional neural network (CNN) in deep learning to directly process meteorological statistical data end-to-end. This method represents the latest progress in the field of agricultural insurance compensation amount prediction. By treating meteorological statistical data as two-dimensional image input, CNN is used to perform convolution, activation, and pooling operations on the data, and then the final compensation amount prediction value is output through the fully connected layer. Specifically, various meteorological indicators (such as precipitation, temperature, wind speed, etc.) during the disaster stage are organized into a set of two-dimensional feature maps. The CNN model can extract spatiotemporal features from these feature maps to realize automatic analysis of meteorological variables. This method skips the specific quantification step of agricultural losses and directly outputs the predicted value of compensation amount, which simplifies the prediction process and improves the prediction speed. However, due to the lack of understanding of the growth mechanism of crops by the convolutional neural network model, the natural recovery characteristics of crops and the cross-influence between various meteorological indicators are ignored, which may lead to large prediction errors under extreme climate conditions. In addition, the prediction performance of the CNN method may be affected in the absence of high-quality training data, and it is not suitable for direct application in agricultural insurance scenarios with high loss risks.

[0006] Combining the above three schemes, the compensation amount prediction scheme for meteorological index insurance has strong pertinence and high prediction accuracy, but its scope of application is narrow, mainly for meteorological index insurance. The scheme of post-disaster yield measurement combined with crop model deduction is highly mature and widely applicable, but its accuracy is highly dependent on manual yield measurement data, which is cumbersome to operate and costly. The scheme based on convolutional neural network significantly simplifies the prediction process and is suitable for application scenarios of rapid claim prediction, but it lacks understanding of the complexity of the relationship between meteorology and crops, and the prediction error may be large when dealing with complex meteorological conditions. The actual application of each scheme can be selected according to the specific agricultural insurance needs and conditions, but as the deep learning model's understanding of the relationship between meteorology and crops continues to deepen, data-driven prediction schemes may play a greater role in the field of agricultural insurance. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a method for predicting the amount of agricultural insurance compensation based on time series external regression, which mainly targets six types of meteorological time series data, and more accurately predicts the agricultural losses that may occur in the planting area and cause claims from historical time series meteorological data, helping agricultural insurance companies to accelerate the compensation process and improve the accuracy of compensation.

[0008] The technical solution of the present invention is as follows:

[0009] A method for predicting the amount of agricultural insurance compensation based on time series external regression considers the prediction problem as a time series external regression task, which predicts the amount of compensation by analyzing the short-term agricultural meteorological time series data before the claim. The prediction method specifically includes the following steps:

[0010] Step 1: Collect meteorological index data, crop planting area data and crop loss compensation data to construct original data;

[0011] Step 2: Process the original data to generate agricultural meteorological time series data;

[0012] Step 3: construct a time series data transformation unit to standardize, fill, crop and stack the agricultural meteorological time series data;

[0013] Step 4: construct additional data encoding units to integrate external information into agricultural meteorological time series data through one-hot encoding;

[0014] Step 5: Build a backbone network, adjust the input and output of the backbone network, and predict the amount of agricultural insurance compensation.

[0015] Furthermore, in step 1, the meteorological index data include temperature, rainfall, humidity, and snow water equivalent; the meteorological index data of the crop planting area within 60 days before the compensation occurs are collected; the crop planting area data contains crop information; the compensation amount for crops damaged by disasters, the year and month when the compensation occurs, the planting area of ​​the insured crops, the crop planting location, and the type of insurance insured are collected to form the crop loss compensation data.

[0016] Furthermore, the specific process of step 2 is as follows:

[0017] Step 2.1, taking each data item in the original data as the main body, time matching is performed on the original data to divide the area for statistical meteorological indicators;

[0018] Step 2.2, combine the crop planting area data to construct the spatiotemporal aligned meteorological data, and use the meteorological index screening method to filter out the non-important meteorological indexes to generate the meteorological index data statistics; the specific process of the meteorological index screening method is: draw a distribution map according to the values ​​of all meteorological indexes, and calculate the standard deviation of all meteorological indexes at the same time. The indexes with concentrated value distribution and small standard deviation are non-important indexes, and the non-important indexes are deleted;

[0019] Step 2.3, using the crop screening method to screen the crop varieties; the specific process of the crop screening method is: counting the number of claims for different crops in the crop loss compensation data, and selecting the crop variety with the largest number of claims;

[0020] Step 2.4: Use bilinear interpolation to match the longitude and latitude of the crop planting area and the longitude and latitude of the meteorological index, and calculate the meteorological index value. The formula is:

[0021]

[0022] Among them, f(x,y) represents the meteorological index value of the meteorological point (x,y) in the planting area, x and y are the longitude and latitude coordinates of the meteorological point (x,y) respectively; (x1,y1), (x2,y1), (x1,y2), (x2,y2) are the positions of the four meteorological points closest to the meteorological point (x,y) in the meteorological data respectively; x1 and x2 are different longitude coordinates; y1 and y2 are different latitude coordinates;

[0023] Step 2.5, perform meteorological index data statistics, and finally calculate the maximum value, minimum value, quartile value, average value, and standard deviation of the meteorological index of each meteorological point every day in the past 60 days at all meteorological points in the crop planting area; at the same time, calculate the inflation rate;

[0024] Step 2.6: Integrate all data to generate agricultural meteorological time series data.

[0025] Furthermore, the specific process of step 3 is as follows:

[0026] Step 3.1: Standardize the agricultural meteorological time series data;

[0027] Define the agricultural meteorological time series data set Z = {z1,z2,…,z n}, the standardized agricultural meteorological time series data set is U = {u1,u2,…,u n}, the standardized formula is as follows:

[0028]

[0029] Among them, u t is the standardized agricultural meteorological time series data at time t; z t is the agricultural meteorological time series data at time t; μ is the mean; σ is the standard deviation; n is the length of the time series; u n is the standardized agricultural meteorological time series data at time n; z n is the agricultural meteorological time series data at time n;

[0030] Step 3.2: Align the agricultural meteorological time series data by padding, cropping, and dimension transformation. The specific process is as follows:

[0031] Step 3.2.1, let O be a 2 ≤Length(U) the largest integer; Length(U) is the length of U; if o 2 = Length(U), then the dimension transformation is performed directly, that is, the element M in the i-th row and j-th column of the aligned agricultural meteorological time series data is set i,j= Z[i×n+j], where i, j = 0, 1, …, n-1; otherwise, proceed to step 3.2.2 to perform cropping and padding, and then perform dimensional transformation;

[0032] Step 3.2.2, if (o+1) 2 -Length(U)>Length(U)-o 2 , then cropping is performed, that is, setting the cropped After cropping, the dimension is transformed, that is, M is set i,j =U[i×o+j], where i,j=0,1,…,o-1; For the 2 -1; otherwise, fill U with 0, that is, the filled After filling, the dimension is transformed, that is, M is set i,j =Z[i×(n+1)+j], i,j=0,1,…,n; U Length(U)-1 , They are located at Length(U)-1 and (o+1) respectively. 2 The data value at

[0033] Step 3.2.3, finally obtain the aligned agricultural meteorological time series data M;

[0034] Step 3.3: Perform channel superposition on M. The formula is:

[0035] N=stack([M r,1 ,M r,2 ,…,M r,E ],axis=0);

[0036] at this time, is the input data after stacking, E is the total number of channels; stack(·) is the stacking; M r,E is the data variable for the r indicator in channel E; axis is the stacking direction.

[0037] Furthermore, in step 4, the external information includes different regions, months, and insurance types; the specific working process of the additional data encoding unit is: the external information is integrated into the agricultural meteorological time series data by one-hot encoding, and then the characteristic dimension of the encoded data is mapped to the same dimension as the number of variables by linear transformation; the calculation formula of one-hot encoding is:

[0038] C one-hot =[δ(C,c1),δ(C,c2),…,δ(C,c h )];

[0039]

[0040] Among them, C one-hot is the data representation after one-hot encoding; δ(·) is the decision unit of one-hot encoding; C is the categorical variable; c h is the number of the corresponding real data in category h.

[0041] Furthermore, in step 5, ShuffleNet v2 is selected to construct the backbone network, the input of the backbone network is adjusted to agricultural meteorological time series data with a size of C×o×o, and the output is adjusted to the compensation amount with a size of 1×1.

[0042] Beneficial technical effects brought by the present invention: The present invention proposes a solution based on time series external regression for the problem of agricultural insurance compensation amount prediction, which can be widely applied to the insurance compensation amount prediction task of various crops. The method relies on the meteorological data integration module to unify and integrate the insurance information of different climate scales and multiple crops. This module not only integrates multiple meteorological indicators of the target area in time series, but also generates the training data required for the prediction model to meet the needs of different crops. In the actual prediction process, the meteorological data integration module can update the data of various insured plots in real time to provide high-quality input for generating prediction results. In order to improve data processing efficiency, the designed time series data conversion unit is used to replace the traditional data encoding method. The unit realizes the automatic transformation of data latitude through filling and clipping methods, and stacks multiple time series data into a multi-channel two-dimensional data format, which is similar to the structure of a multi-dimensional image, thereby simplifying the input format. In addition, the method no longer relies on the traditional sequence decomposition encoding method, but instead hands over the trend, seasonality, periodicity and other feature recognition tasks to the subsequent deep learning backbone network for processing. Doing so not only avoids the potential impact of encoding inconsistency on model performance, but also greatly reduces the computational cost of the encoding stage. The additional data encoding unit further enhances the model's predictive ability, seamlessly integrating external information that may be related to the compensation amount into the original data, allowing the model to capture more key features related to the claim. The backbone network stage uses the optimized ShuffleNet v2 structure to specifically model the intrinsic correlation of time series data and generate the final claim prediction value. The invention deeply explores the impact of meteorological data on crop losses, establishes a feasible compensation amount prediction model for various types of agricultural insurance, and improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flow chart of the method for predicting the amount of agricultural insurance compensation based on time series external regression.

[0044] Figure 2 Schematic diagram of the channel superposition process of the present invention.

[0045] Figure 3 Schematic diagram of the one-hot encoding process of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0047] The present invention proposes an innovative method for predicting the amount of agricultural insurance compensation. The prediction problem is regarded as a time series external regression task, which predicts the amount of compensation by analyzing the short-term agricultural meteorological time series data before the claim. Agricultural meteorological time series data are constructed based on temporal meteorological indicators, cultivated land plot data and crop loss compensation data. A simple and effective solution is proposed using the existing ShuffleNet V2 model. Through the time series data transformation unit and the additional data encoding unit, the processed data is sent to the backbone network for prediction. This method achieves performance that is ahead of specially designed time series analysis algorithms while significantly reducing the computational complexity.

[0048] like Figure 1 As shown in FIG. 1 , a method for predicting agricultural insurance compensation amount based on time series external regression includes the following steps:

[0049] Step 1: Collect meteorological indicators, crop planting areas and crop loss compensation data to construct original data;

[0050] Agricultural meteorological indicators such as temperature, rainfall, humidity, and snow water equivalent have an important impact on the growth of crops. These meteorological indicator data cover a wide range and need to be selectively processed in a targeted manner. At the same time, due to the wide distribution of crop planting areas and the wide variety of crops planted, different crops have different crop mechanisms, and the losses caused by different meteorological changes are also different. Therefore, it is necessary to integrate high-resolution meteorological indicator data and crop loss compensation data in a targeted manner. Through data integration, historical meteorological indicator data of specific plots can be effectively extracted, and grouped and aggregated using specific aggregation algorithms, thereby generating reasonable and effective agricultural meteorological time series data.

[0051] As the ultimate goal of prediction, the spatiotemporal granularity of crop loss compensation data determines the upper limit of the overall refinement of the data. When acquiring data, we collect information such as the amount of compensation for specific crops damaged by disasters, the year and month of compensation, the planting area of ​​insured crops, the location of crop planting (county), and the type of insurance insured, and then generate crop loss compensation data.

[0052] For meteorological indicator data, the time series statistics are limited to the 60 days before the compensation occurs, and only the meteorological indicator data of specific crop planting areas are counted, rather than the meteorological indicator data of the entire county, so as to generate the required meteorological indicator data.

[0053] Step 2: Process the original data to generate agricultural meteorological time series data;

[0054] The obtained crop loss compensation data, meteorological index data, and crop planting area data contain rich crop information, meteorological information, and compensation information, but their time, spatial dimensions, and data items are not aligned, and fusion matching algorithms and data cleaning methods are needed to ensure the availability of the data. The crop screening method is used to control the crop varieties included in the test, and the data is further refined according to the longitude and latitude matching algorithm. Given that the data time span is too large, the amount of crop loss compensation may not truly reflect the corresponding relationship. Therefore, the inflation rate calculation is included in the construction process. For each agricultural insurance compensation amount, it is converted into the current currency value according to the inflation rate in the year when the compensation occurs. Finally, the agricultural meteorological time series data is generated by fusion. The specific process is:

[0055] Step 2.1: Taking each data item in the original data as the main body, perform time matching on the original data and divide the areas that can be used for statistical meteorological indicators.

[0056] Step 2.2: Combine crop planting area data to construct time-space aligned meteorological data, and use meteorological index screening methods to filter out non-important meteorological indicators to generate meteorological index data statistics. Draw a distribution map based on the values ​​of all meteorological indicators, and calculate the standard deviation of all meteorological indicators. Indicators with relatively concentrated value distribution and extremely small standard deviation are non-important indicators.

[0057] Step 2.3: Use crop screening methods to control the crop varieties included in the test; count the number of claims for different crops in the crop loss compensation data, and select the crop varieties with the largest number of claims to help model training;

[0058] Step 2.4: Use bilinear interpolation to match the longitude and latitude of the crop planting area selected in the county with the longitude and latitude of the meteorological index, and calculate the meteorological index value. The calculation formula for the meteorological index value of each meteorological point in the crop planting area is as follows:

[0059]

[0060] Among them, f(x,y) represents the meteorological index value of the meteorological point (x,y) in the planting area, x and y are the longitude and latitude coordinates of the meteorological point (x,y), respectively; (x1,y1), (x2,y1), (x1,y2), (x2,y2) are the positions of the four meteorological points closest to the meteorological point (x,y) in the meteorological data, respectively; x1 and x2 are different longitude coordinates; y1 and y2 are different latitude coordinates.

[0061] Step 2.5, conduct meteorological index data statistics, and finally calculate the maximum value, minimum value, quartile value, average value, and standard deviation of the meteorological index of each meteorological point every day in the past 60 days in all meteorological points in the crop planting area. Simultaneously, use existing technology to calculate the inflation rate;

[0062] Step 2.6: Integrate all data to generate agricultural meteorological time series data;

[0063] Meteorological data integration integrates raw data precisely in two dimensions, and builds high-quality data tables at the smallest granularity (such as plot-date) by aligning spatial and temporal dimensions. Based on the longitude and latitude information of the planting area, meteorological data is aligned to specific plots through bilinear interpolation to ensure that meteorological data from different sources have consistent spatial resolution. Using the plot's crop growth cycle and analysis target time range, the temporal dimension of meteorological data is aligned with the plot records to ensure that each plot has continuous meteorological indicators over the full time range.

[0064] Step 3: Construct a time series data transformation unit, standardize the agricultural meteorological time series data, align the agricultural meteorological time series data by padding and clipping, stack the data of different variables into multiple channels, and construct a format similar to a multidimensional channel. This method omits the traditional time series data encoding stage, effectively reduces the potential impact of encoding inconsistency on model performance, significantly reduces the computational cost of this stage, and improves the final accuracy of the agricultural insurance compensation amount.

[0065] Agricultural meteorological time series data require special attention due to its strong time dependence. The order of data points in the time series of different meteorological data directly affects the accuracy of the prediction results, and there is often a short-term or long-term correlation between data points of different dimensions. Therefore, when processing time series data, its time order and the correspondence between the time points of each variable must be retained to ensure that the data processing process is completely corresponding to the original data. A simple and efficient data transformation unit first standardizes the data. After the data is standardized, in order to retain the characteristics of the original sequence data as much as possible, the time series data is aligned by padding or cropping, and the data of different variables are stacked into different channels to construct a data format similar to a multidimensional channel. By omitting the encoding stage, the negative impact of encoding inconsistency on model performance is reduced, and the computational burden of this stage is significantly reduced.

[0066] The time series data transformation unit operates on the basis of the original integrated meteorological data, making the data form meet the needs of the backbone network, while retaining the correlation between different variables, which is convenient for the backbone network to model long-term and short-term relationships. The specific process is:

[0067] Step 3.1: Standardize the agricultural meteorological time series data;

[0068] For the agricultural meteorological time series data set Z = {z1,z2,…,z n The purpose of standardization is to adjust the mean of the data to 0 and the variance to 1, so as to improve the stability and convergence speed of the subsequent model and improve the generalization ability of the model. The standardized agricultural meteorological time series data set is U = {u1,u2,…,u n}, the standardized mathematical expression is as follows:

[0069]

[0070] Among them, u t is the standardized agricultural meteorological time series data at time t; z t is the agricultural meteorological time series data at time t; μ is the mean; σ is the standard deviation; n is the length of the time series; u n is the standardized agricultural meteorological time series data at time n; z n is the agricultural meteorological time series data at time n.

[0071] Step 3.2, trimming or filling by judging the length of the time series, and aligning the agricultural meteorological time series data by padding, trimming, and dimension transformation;

[0072] The agricultural meteorological time series data after standardization is still one-dimensional time series data, which needs to be converted into stacked two-dimensional time series data for the convenience of subsequent model input. In order to retain the characteristics of the original sequence data as much as possible, the influence of the sequence length on the conversion result is fully considered during the two-dimensional conversion. If the length of the converted sequence is slightly larger than the original sequence length, the data is padded with zero values; if it is slightly smaller than the original length, the excess part will be trimmed. The specific method is to calculate the maximum square root o that is less than or equal to U, and then compare o 2 and (o+1) 2 The difference between the length n of the original time series is used to determine whether to crop the original data or fill it with zero values, and finally transform the data dimension into a two-dimensional tensor;

[0073] The specific process of aligning agricultural meteorological time series data is as follows:

[0074] Step 3.2.1, let o be a 2 ≤Length(U) the largest integer; Length(U) is the length of U; if o 2= Length(U), then the dimension transformation is performed directly, that is, the element M in the i-th row and j-th column of the aligned agricultural meteorological time series data is set i,j = Z[i×n+j], where i, j = 0, 1, …, n-1; otherwise, proceed to step 3.2.2 to perform cropping and padding, and then perform dimensional transformation;

[0075] Step 3.2.2, if (o+1) 2 -Length(U)>Length(U)-o 2 , then cropping is performed, that is, setting the cropped After cropping, the dimension is transformed, that is, M is set i,j =U[i×o+j], where i,j=0,1,…,o-1; For the 2 -1; otherwise, fill U with 0, that is, the filled After filling, the dimension is transformed, that is, M is set i,j =Z[i×(n+1)+j], i,j=0,1,…,n; U Length(U)-1 , They are located at Length(U)-1 and (o+1) respectively. 2 The data value at

[0076] Step 3.2.3, finally obtain the aligned agricultural meteorological time series data M;

[0077] Step 3.3, superimpose different indicator variables in the meteorological indicator data into multiple channels;

[0078] The impact of meteorological index data on crops is caused by the joint influence of multiple related indicators. Therefore, the meteorological index is essentially a multivariate indicator composed of several indicator variables. When processing data, the temporal and spatial relationships of different indicator variables are matched, while retaining the internal relationship of a single indicator. Figure 2 As shown in , for multivariate indicators, variable channels are first separated to form several single indicator variables, and then each indicator is standardized and dimensionally transformed in the same way as the single variable indicator. Finally, in order to maintain the time correspondence between different indicator variables, the M corresponding to each indicator variable is stacked together to form multiple channels, thereby constructing a data format similar to a multidimensional image, as shown in Figure 2 shown.

[0079] The agricultural meteorological time series data M after all channels are aligned are stacked in the channel dimension to obtain the stacked input data:

[0080] N=stack([M r,1 ,Mr,2 ,…,M r,E ],axis=0);

[0081] at this time, is the input data after stacking, E is the total number of channels; stack(·) is the stacking; M r,E is the data variable for the r indicator in channel E; axis is the stacking direction;

[0082] at this time The corresponding channel This processing method helps the subsequent learning of the convolutional neural network backbone network, and ensures the preservation of the time series correspondence between multiple variables from the data level. Unlike conventional low-dimensional mapping methods, conventional methods map low-dimensional multivariate data to high-dimensional space, which not only increases the computational complexity of the model but also may introduce the risk of inaccurate learning. By directly stacking different variables into channels, potential errors are avoided, making the model more reliable when processing multivariate time series data. This enables the model to fully utilize the advantages of convolutional neural networks in processing high-dimensional data, while ensuring that the time series relationship between multiple variables does not change during the data conversion process.

[0083] Step 4. Aiming at the characteristics of the task of predicting the amount of agricultural insurance compensation by time series meteorological indicators, the present invention constructs an additional data encoding unit, integrates external information of different types and units into the time series data by one-hot encoding, and then uses linear transformation to map the characteristic dimension of the encoded data to the same dimension as the number of variables, thereby enhancing the sensitivity of the model to external information and improving the model's ability to extract features of time series data;

[0084] Traditional time series encoding methods mainly encode values ​​and positions. In the scenario of predicting agricultural insurance compensation amounts based on meteorological time series data, external information is not only related to time, but also closely related to the external environment. In the task of predicting agricultural insurance compensation amounts based on time series meteorological indicators, external information such as different regions, months, and insurance types also plays an important role in the prediction results. Integrate external information into the feature extraction of time series data and use one-hot encoding to convert it into a binary vector, expressed as:

[0085] C one-hot =[δ(C,c1),δ(C,c2),…,δ(C,c h )];

[0086]

[0087] Among them, C one-hotis the data representation after one-hot encoding; δ(·) is the decision unit of one-hot encoding; C is the categorical variable; c h is the number of the corresponding real data in category h.

[0088] like Figure 3 As shown, the present invention uses linear transformation on the additional information after one-hot encoding to obtain superimposed data, thereby realizing the mapping of the encoded feature dimension to the same dimension as the number of time series variables. The encoded information is attached to the data at each time point according to the time characteristics, thereby improving the model's ability to analyze time series data. Under this processing method, the overall dimension of the input data remains unchanged, but the values ​​of each variable at each time point change due to the addition of additional information.

[0089] Finally, the data processed by the additional data coding unit are fused and integrated into the agricultural meteorological time series data processed in step 3 to serve as the input data for subsequent models.

[0090] Step 5: Build a backbone network, adjust the input and output of the backbone network, and predict the amount of agricultural insurance compensation;

[0091] The backbone network is the core of the entire framework, responsible for extracting and encoding the features of meteorological time series data to achieve accurate prediction of the amount of agricultural insurance compensation. In the selection of the backbone network, the characteristics of the input data, namely the multidimensionality and complexity of the agricultural meteorological time series data, must be taken into account to ensure that the network architecture can capture the subtle changes in the time series data and the association between multiple variables. The main goal of the present invention is to predict the final amount of compensation based on the agricultural meteorological time series data. The previous part has converted the multivariate one-dimensional time series data into stacked two-dimensional data. The backbone network based on ShuffleNet v2 is selected, and the application framework network can be obtained by making corresponding adjustments to the input and output. ShuffleNet v2 is a lightweight convolutional neural network architecture designed for efficient image classification tasks on mobile and embedded devices. It is further optimized on the basis of ShuffleNet v1. ShuffleNet v1 uses grouped convolution to accelerate network operations by channel shuffling, while ShuffleNet v2 proposes a channel separation operation, which reuses features while accelerating the network, achieving better results. The typical input of the ShuffleNet v2 model is an image of size 3×28×28, and the output is a classification score of size k×1, where k is the number of image types. In the present invention, the input is stacked two-dimensional agricultural meteorological time series data of size C×o×o, and the output is the compensation amount of size 1×1. In order to meet this demand, only the input layer of the model needs to be adjusted to accept C channel data input, and the output layer needs to be adjusted to 1 channel, without changing the intermediate structure of the model.

[0092] The present invention adjusts the input and output dimensions of the original ShuffleNet v2 lightweight network so that the network can adapt to the task of predicting the amount of agricultural insurance compensation. At the same time, with the help of the network's special convolution design and channel design, especially its group convolution and channel shuffling mechanism, the accuracy of the model prediction is guaranteed while significantly reducing the amount of calculation. The present invention can effectively improve the model's processing efficiency for multi-channel meteorological data, so that it also has good adaptability in mobile and embedded scenarios. The lightweight ShuffleNetv2 is selected as the backbone network of the framework, while minimizing the model's calculation complexity and parameter quantity, and excellent prediction performance is obtained. The technology of predicting the amount of agricultural insurance compensation by the time series external regression framework based on ShuffleNet v2 can help agricultural insurance companies to conduct early confirmation of insurance claims, improve the timeliness and accuracy of insurance payments.

[0093] In order to prove the feasibility and superiority of the present invention, the following comparative experiment is given.

[0094] The meteorological data integration technology introduced in the invention was used to integrate and construct the soybean insurance planting claim data of 645 counties in seven states in the United States from 2008 to 2022. The construction finally generated 133,424 data items, covering the agricultural insurance claims caused by drought and waterlogging disasters of soybeans and their corresponding meteorological statistical indicators within 60 days. The data items were divided into a ratio of 7:3, 70% as a training set and 30% as a test set. Under the specified training method, the mean absolute error MAE of the test results was 57.02, and the root mean square error RMSE was 91.37. It is proved that the model has a significant effect in predicting the amount of agricultural insurance claims based on time series meteorological data.

[0095] At the same time, compared with other time series prediction models, the comparison results are shown in Table 1:

[0096] Table 1 Comparison of results of different models

[0097]

[0098]

[0099] Transformer is a deep learning model architecture, which was first proposed in 2017. It is mainly used to process sequence data, especially in the field of natural language processing. Compared with traditional recurrent neural networks and long short-term memory networks, it improves computational efficiency and the ability to capture global dependencies. Informer is a deep learning model for long time series prediction problems. It is based on the Transformer architecture and is specially designed to handle long-term dependencies and large-scale data sets. Pyraformer is a deep learning model for time series prediction. It combines the advantages of Transformer and recurrent neural networks to improve the accuracy and computational efficiency of long time series prediction. FEDformer is a deep learning model for time series prediction problems. It combines the advantages of Fourier transform and Transformer architecture to more efficiently process time series data, especially in the case of long-term dependencies. TimesNet is a deep learning model designed for time series data. It is often used to predict and analyze time series data. It combines convolutional neural networks and self-attention mechanisms to handle long sequence dependency problems and capture time series features by using these techniques. LSTM is a long short-term memory network.

[0100] It can be seen from Table 1 that the computational complexity and parameter quantity of the model of the present invention are the lowest, so the model of the present invention has a significant computational complexity advantage.

[0101] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting agricultural insurance compensation amount based on time series external regression, characterized in that: The prediction problem is regarded as a time series external regression task, which predicts the compensation amount by analyzing the short-term agricultural meteorological time series data before the claim. The prediction method specifically includes the following steps: Step 1: Collect meteorological index data, crop planting area data and crop loss compensation data to construct original data; Step 2: Process the original data to generate agricultural meteorological time series data; Step 3: construct a time series data transformation unit to standardize, fill, crop and stack the agricultural meteorological time series data; Step 4: construct an additional data encoding unit to integrate external information into the agricultural meteorological time series data through one-hot encoding; Step 5: Build a backbone network, adjust the input and output of the backbone network, and predict the amount of agricultural insurance compensation.

2. The agricultural insurance compensation amount prediction method based on time series external regression according to claim 1 is characterized in that: In step 1, meteorological index data include temperature, rainfall, humidity, and snow water equivalent; meteorological index data of crop planting areas within 60 days before compensation occurs are collected; crop planting area data include crop information; the amount of compensation for crops damaged by disasters, the year and month of compensation, the planting area of ​​insured crops, the crop planting location, and the type of insurance insured are collected to form crop loss compensation data.

3. The agricultural insurance compensation amount prediction method based on time series external regression according to claim 1 is characterized in that: The specific process of step 2 is: Step 2.1, taking each data item in the original data as the main body, time matching the original data is performed to divide the area for statistical meteorological indicators; Step 2.2, combine the crop planting area data to construct the spatiotemporal aligned meteorological data, and use the meteorological index screening method to filter out the non-important meteorological indexes to generate the meteorological index data statistics; the specific process of the meteorological index screening method is: draw a distribution map according to the values ​​of all meteorological indexes, and calculate the standard deviation of all meteorological indexes at the same time. The indexes with concentrated value distribution and small standard deviation are non-important indexes, and the non-important indexes are deleted; Step 2.3, using the crop screening method to screen the crop varieties; the specific process of the crop screening method is: counting the number of claims for different crops in the crop loss compensation data, and selecting the crop variety with the largest number of claims; Step 2.4: Use bilinear interpolation to match the longitude and latitude of the crop planting area and the longitude and latitude of the meteorological index, and calculate the meteorological index value. The formula is: Among them, f(x,y) represents the meteorological index value of the meteorological point (x,y) in the planting area, x and y are the longitude and latitude coordinates of the meteorological point (x,y) respectively; (x1,y1), (x2,y1), (x1,y2), (x2,y2) are the positions of the four meteorological points closest to the meteorological point (x,y) in the meteorological data respectively; x1 and x2 are different longitude coordinates; y1 and y2 are different latitude coordinates; Step 2.5, perform meteorological index data statistics, and finally calculate the maximum value, minimum value, quartile value, average value, and standard deviation of the meteorological index of each meteorological point every day in the past 60 days at all meteorological points in the crop planting area; at the same time, calculate the inflation rate; Step 2.6: Integrate all data to generate agricultural meteorological time series data.

4. The method for predicting agricultural insurance compensation amount based on time series external regression according to claim 1 is characterized in that: The specific process of step 3 is as follows: Step 3.1: Standardize the agricultural meteorological time series data; Define the agricultural meteorological time series data set Z = {z1,z2,…,z n }, the standardized agricultural meteorological time series data set is U = {u1,u2,…,u n }, the standardized formula is as follows: Among them, u t is the standardized agricultural meteorological time series data at time t; z t is the agricultural meteorological time series data at time t; μ is the mean; σ is the standard deviation; n is the length of the time series; u n is the standardized agricultural meteorological time series data at time n; z n is the agricultural meteorological time series data at time n; Step 3.2: Align the agricultural meteorological time series data by padding, cropping, and dimension transformation. The specific process is as follows: Step 3.2.1, let o be a 2 ≤Length(U) the largest integer; Length(U) is the length of U; if o 2 = Length(U), then the dimension transformation is performed directly, that is, the element M in the i-th row and j-th column of the aligned agricultural meteorological time series data is set i,j = Z[i×n+j], where i, j = 0, 1, …, n-1; otherwise, proceed to step 3.2.2 to perform cropping and padding, and then perform dimensional transformation; Step 3.2.2, if (O+1) 2 -Length(U)>Length(U)-o 2 , then cropping is performed, that is, setting the cropped After cropping, the dimension is transformed, that is, M is set i,j = U[i×o+j], where i,j=0,1,…,O-1; For the 2 -1; otherwise, fill U with 0, that is, the filled After filling, the dimension is transformed, that is, m is set i,j =Z[i×(n+1)+j], i,j=0,1,…,n; U Length(U)-1 , They are located at Length(U)-1 and (o+1) respectively. 2 The data value at Step 3.2.3, finally obtain the aligned agricultural meteorological time series data M; Step 3.3: Perform channel superposition on M. The formula is: N=stack([M r,1 ,M r,2 ,…,M r,E ],axis=0); at this time, is the input data after stacking, E is the total number of channels; stack(·) is the stacking; M r,E is the data variable for the r indicator in channel E; axis is the stacking direction.

5. The method for predicting agricultural insurance compensation amount based on time series external regression according to claim 4 is characterized in that: In step 4, the external information includes different regions, months, and insurance types; the specific working process of the additional data encoding unit is: the external information is integrated into the agricultural meteorological time series data by one-hot encoding, and then the characteristic dimension of the encoded data is mapped to the same dimension as the number of variables by linear transformation; the calculation formula of one-hot encoding is: C one-hot =[δ(C,c1),δ(C,c2),…,δ(C,c h )]; Among them, C one-hot is the data representation after one-hot encoding; δ(·) is the decision unit of one-hot encoding; C is the categorical variable; c h is the number of the corresponding real data in category h.

6. The method for predicting agricultural insurance compensation amount based on time series external regression according to claim 5 is characterized in that: In step 5, ShuffleNet v2 is selected to construct a backbone network, the input of the backbone network is adjusted to agricultural meteorological time series data with a size of C×o×o, and the output is adjusted to the compensation amount with a size of 1×1.

Images