A cascaded feature extraction method for agricultural product price prediction
By employing a cascaded feature extraction method, combined with KNN-MI, XGBoost, and multi-scale wavelet decomposition, the problem of processing high-dimensional and nonlinear features in agricultural product price prediction was solved, achieving higher prediction accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- COLLEGE OF MOBILE TELECOMM CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-05
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies for agricultural product price forecasting suffer from limitations of traditional time series models, reliance on low-quality features in single-stage forecasting models, and a lack of systematicity in existing feature engineering methods, resulting in insufficient forecast accuracy and stability, and difficulty in handling high-dimensional, nonlinear, and non-stationary features.
A cascaded feature extraction method is adopted, including KNN-MI to capture nonlinear dependencies, XGBoost to evaluate feature importance, and multi-scale wavelet decomposition to enhance features. Feature selection and enhancement are performed through a three-level cascaded architecture, and prediction is performed in conjunction with a BiLSTM model.
It significantly improves the accuracy and robustness of agricultural product price forecasting, solves the problem of high-dimensional data processing, and enhances the interpretability of feature selection and the stability of the prediction model.
Smart Images
Figure CN122155801A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and agricultural economic analysis technology, specifically a cascaded feature extraction method for agricultural product price prediction. Background Technology
[0002] Agricultural product price values are a core indicator reflecting market price fluctuations, and their standardized forecasting is of great significance for national food security, market stability, and agricultural economic decision-making. With the increasing volume and dimensionality of agricultural product price data, price forecasting research has shifted from traditional econometric models to data-driven machine learning methods. However, existing technologies face the following prominent bottlenecks in practical applications:
[0003] First, there are limitations to traditional time series forecasting models. Early research primarily relied on econometric models and classical time series models. These methods, based on rigorous theoretical assumptions and artificially defined equations, struggle to handle agricultural product price data characterized by nonlinearity and non-stationarity. Moreover, traditional models have limited capacity to incorporate multi-source, high-dimensional external influencing factors, and even when they are included, multicollinearity often leads to estimation distortion. Therefore, these methods are insufficient in explanatory power and forecasting accuracy when dealing with the complex market environment of today.
[0004] Secondly, there are inherent flaws in the direct application of single-stage prediction models. To overcome the limitations of the linear assumptions of traditional models, support vector machines, random forests, gradient boosting decision trees, and deep learning models have been widely used in agricultural product price prediction in recent years. Although these models have significant advantages in nonlinear fitting capabilities, a crucial premise that is often overlooked is that model performance is highly dependent on the quality of input features. In practice, researchers often directly input raw, multi-source features without systematic processing into the model. This leads to two major problems: First, the curse of dimensionality and the risk of overfitting. Features related to agricultural product prices can have tens or even hundreds of dimensions, including a large number of redundant, noisy, or features with very weak correlation to the prediction target. Direct use of these features will drastically increase the model's complexity, causing it to tend to memorize noise in the data rather than learn general patterns, resulting in poor generalization ability on unknown data. Second, even with an advanced model structure, low-quality feature input cannot produce high-quality prediction output. Redundancy and noise in the features will mask the real key driving signals, making it difficult for the model to converge to the optimal state.
[0005] Third, existing feature engineering methods lack systematicity and depth. After recognizing the importance of feature quality, some studies have begun to introduce feature selection or construction techniques. However, existing methods are relatively simplistic. Traditional approaches only use correlation coefficients, mutual information, and other methods for one-time screening. These methods are computationally efficient, but the evaluation criteria are independent of the subsequent prediction model. The selected features may be relevant to the target when viewed individually, but when combined, they may be informationally redundant, and the interaction effects between features cannot be evaluated.
[0006] A search revealed application publication number CN114493697A, which discloses a method, system, storage medium, and electronic device for predicting agricultural product prices. The method includes: acquiring historical agricultural product price data; normalizing the price data and using LASSO regression to select the optimal input variables; using cross-validation to select the optimal hyperparameters to obtain the best model for different quantiles; training the selected best model using historical agricultural product price data to obtain a trained model; acquiring current agricultural product price data; normalizing the current agricultural product price data and substituting it into the trained model for prediction to obtain k quantiles; performing kernel density estimation based on the k quantiles to obtain the probability density distribution of the predicted agricultural product price; and outputting the probability density prediction distribution result. Therefore, the agricultural product price prediction method of this invention can perform multi-dimensional prediction of agricultural product prices and provide more information, making the prediction results more reliable.
[0007] The aforementioned invention relies on LASSO regression for variable selection, which has significant technical limitations: First, LASSO regression is based on a linear assumption and can only capture the linear relationship between features and agricultural product prices, failing to identify the nonlinear and non-monotonic dependencies that are common in the actual market, and easily missing key nonlinear features; Second, this feature selection method is singular and lacks systematicity, making it difficult to effectively filter noise and weakly correlated components when faced with high-dimensional data related to agricultural product prices, and also failing to uncover the interaction effects between features; Third, the one-time variable selection model lacks progressive verification and optimization, resulting in poor interpretability of feature selection and difficulty in quantifying the actual contribution of selected features to predictive performance. To address the aforementioned shortcomings, this invention specifically constructs a cascaded feature extraction method: it captures nonlinear dependencies through KNN-MI to overcome the limitations of linear screening; it utilizes XGBoost for gain-based feature importance assessment to systematically eliminate redundant and noisy features and uncover interaction effects, solving the challenges of high-dimensional data processing; and it strengthens the representation of temporal features through multi-scale wavelet decomposition, coupled with systematic ablation experiments to quantify the contribution of each processing stage, significantly improving the interpretability and rigor of feature selection, thus comprehensively breaking through the bottlenecks of existing technologies. Summary of the Invention
[0008] This invention aims to solve the problems of the prior art mentioned above. It proposes a cascaded feature extraction method for agricultural product price forecasting. The technical solution of this invention is as follows:
[0009] A cascaded feature extraction method for agricultural product price forecasting includes the following steps:
[0010] S1: Acquire multi-source heterogeneous data on agricultural products and preprocess them to form an original feature set;
[0011] S2: The original feature set is subjected to feature extraction and enhancement using a three-level cascaded architecture to obtain an enhanced feature matrix;
[0012] S3: Construct a prediction model based on the enhanced feature matrix, and verify the effectiveness of the cascaded architecture through ablation experiments;
[0013] S4: Use the trained prediction model to predict agricultural product prices.
[0014] Furthermore, step S2 employs a three-level cascaded architecture to extract and enhance features from the original feature set, obtaining an enhanced feature matrix, specifically including:
[0015] S21: Based on the K-nearest neighbor mutual information estimation method, calculate the mutual information value between each feature in the original feature set and the target price index, and obtain a primary feature subset according to a preset threshold or ranking. ;
[0016] S22: Using the primary feature subset Using the input as input, an XGBoost regression model is trained, and features are extracted based on the importance ranking of split gain. The top-ranked features are selected to form the core feature set. ;
[0017] S23: For the core feature set Wavelet decomposition is performed on the time series features to extract multi-scale components as derived features, and then combined with... Non-temporal features are merged to form an enhanced feature matrix. .
[0018] Furthermore, in step S21, a nonparametric estimation method based on K-nearest neighbors is used to calculate the mutual information value. The estimation formula is:
[0019] ;
[0020] in, For the sample size, For nearest neighbor parameters, This is the Digamma function. The first Each sample in features Univariate space, target variable The corresponding univariate space Number of nearest neighbor samples. Is it the Digamma function in and The function value at that location. It is any one of the influencing factor features in the original feature set for agricultural product price prediction. Target agricultural product prices.
[0021] Furthermore, in step S22, the feature importance of the XGBoost model is obtained by summing the gain values of the feature at all split nodes in each regression tree, and the calculation formula is as follows:
[0022] ;
[0023] in, The total number of trees, Indicates the first Using features in a tree The set of nodes to be split. For the first Tree species characteristics The specific identifier of the split node when executing node splitting; For the first Tree species characteristics At the split node The split gain value corresponding to the split is executed at that point.
[0024] Furthermore, in step S23, the time series features are decomposed by discrete wavelet transform to obtain an approximate coefficient sequence and a detail coefficient sequence, which respectively represent the long-term trend and the multi-scale fluctuation mode, and these components are added to the feature matrix as new features.
[0025] Furthermore, in step S3, the prediction model employs a bidirectional long short-term memory network (BiLSTM) and is trained using the mean squared error loss function.
[0026] Furthermore, the ablation experiment in step S3 includes the following control group:
[0027] Control group A: Using the original feature set;
[0028] Control group B: Using a feature subset selected only by KNN-MI ;
[0029] Control group C: Feature subset selected using KNN-MI and XGBoost ;
[0030] Experimental group D: The enhanced feature matrix F_enhanced after full three-level processing was used; the effectiveness of each level of feature processing was verified by comparing the prediction performance of each control group on the test set.
[0031] Furthermore, the performance metrics include root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²), and the statistical significance of the difference in prediction accuracy is verified using the Diebold-Mariano test.
[0032] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any one of the above.
[0033] The advantages and beneficial effects of this invention are as follows:
[0034] This invention proposes a cascaded feature extraction and enhancement method for agricultural product price forecasting. This method systematically addresses four major pain points in agricultural product price forecasting: high data dimensionality, significant noise interference, complex time-series patterns, and non-linear dependencies among influencing factors. It effectively improves the accuracy of agricultural product price forecasting. Specific advantages and benefits are as follows:
[0035] 1. A systematic cascaded feature extraction and enhancement architecture is proposed (corresponding to claims 1 and 2). A three-level cascaded processing framework is proposed, from filtering coarse screening to embedded fine screening and then to temporal enhancement construction. This method solves the problems of traditional feature engineering being fragmented, isolated, dependent on expert experience, and lacking a systematic approach to problem-solving. It forms a complete feature processing paradigm for complex prediction problems in specific domains, which is not a combination of conventional or known methods used by technicians to solve similar problems.
[0036] 2. Improved accuracy and robustness of the prediction model (corresponding to claims 2, 3, 4, 5, and 6). The proposed method performs dimensionality reduction and noise reduction on the data, improving the signal-to-noise ratio and solving the problems of model accuracy bottleneck, overfitting, and poor stability caused by low-quality input features (redundancy, noise, and pattern aliasing). It aims to ensure the completeness of feature correlation evaluation and is not a conventional technique.
[0037] 3. A rigorous ablation experiment was designed, innovatively incorporating it as an essential component of the method (corresponding to claims 7 and 8). Through strict control of variable comparisons, the independent contribution of each treatment stage to the final performance was quantitatively verified, scientifically demonstrating the necessity and optimization of the framework design. This addresses the lack of empirical evidence on the contribution of internal modules found in many patents, greatly enhancing the persuasiveness and rigor of the solution, which is not typical content in patent applications. Attached Figure Description
[0038] Figure 1 This is a flowchart of the overall method provided by a preferred embodiment of the present invention.
[0039] Figure 2 This is a detailed flowchart of the three-level cascaded feature extraction (S2 step), which is the core of this invention.
[0040] Figure 3 This is a schematic diagram of the logic of the ablation experimental control group designed in one embodiment of the present invention. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.
[0042] The technical solution of the present invention to solve the above-mentioned technical problems is:
[0043] This invention provides a method for predicting agricultural product price indices based on multi-cascade feature extraction, including:
[0044] S1: Multi-source data acquisition and preprocessing.
[0045] S2: Construct a cascaded feature extraction and enhancement method.
[0046] S3: Predictive model construction and ablation experiment.
[0047] S4: Model evaluation and prediction applications.
[0048] Specifically, step S2 includes the following sub-steps:
[0049] Step S21: Calculate the correlation strength between each feature in the original feature set and the target price index, and select a primary feature subset based on a preset threshold. This achieves efficient dimensionality reduction for the first time.
[0050] This patent constructs a KNN-MI feature filter.
[0051] The core advantage of mutual information (MI) lies in its ability to capture non-linear, non-monotonic dependencies between variables. MI measures the relationship between agricultural commodity prices Y and feature X.
[0052] ;
[0053] Since the actual data sample is limited, this invention employs a nonparametric estimation algorithm based on K-nearest neighbors to estimate the mutual information value from the sample data. This is used for subsequent feature selection. Its estimation formula can be expressed as:
[0054] ;
[0055] in, For the sample size, For nearest neighbor parameters, This is the Digamma function. The first Each sample in features Univariate space, target variable The corresponding univariate space Number of nearest neighbor samples. Is it the Digamma function in and The function value at that location. It is any one of the influencing factor features in the original feature set for agricultural product price prediction. The target agricultural product price. Based on a preset threshold or retaining the top-ranked... The rules for each feature are used to complete the initial screening and obtain the feature set F1.
[0056] Step S22: Using the primary feature subset For input, target price index To output the model, an XGBoost regression model is trained. After training, the model's feature importance is ranked based on gain, which directly measures the contribution of each feature to improving the model's prediction accuracy.
[0057] 1. Objective function of XGBoost model
[0058] XGBoost learns using an additive model. In the t-th iteration, its objective function consists of a loss function and a regularization term.
[0059] ;
[0060] in, It is a differentiable convex loss function (such as mean square error). It is the first Before each sample The predicted value of the wheel, It is the tree model to be learned in the t-th round. It is a regularization term that controls the complexity of the model, defined as:
[0061] ;
[0062] in, It is a tree The number of leaf nodes, It is the first The weight of each leaf node and It is a hyperparameter that controls the complexity and weight of the tree.
[0063] 2. Feature Importance (Gain) Calculation
[0064] When constructing each regression tree, for any candidate feature at the current node I... Split point Calculate the split gain (Gain). This gain measures the reduction in the loss function resulting from this split, and its calculation formula is as follows:
[0065] ;
[0066] in:
[0067] and These are the sample sets of the left and right child nodes after the split, respectively.
[0068] It is a loss function The first-order gradient (residual) of the current predicted value.
[0069] It is a loss function The second gradient with respect to the current predicted value.
[0070] The algorithm will traverse all possible features and split points, and select the combination with the largest Gain value to perform the actual split.
[0071] 3. Cumulative and Filtering of Global Importance
[0072] feature Global importance score It is defined as the sum of the Gain values corresponding to all splits in all trees when it is used as a split feature:
[0073] ;
[0074] in, The total number of trees, Indicates the first Using features in a tree The set of nodes to be split. For the first Tree species characteristics The specific identifier of the split node when executing node splitting; For the first Tree species characteristics At the split node The split gain value corresponding to the split is executed at that point. After calculating the importance of all features, they are sorted in descending order of score, and the top-ranked features are selected. The features constitute the core feature set :
[0075] ,in ;
[0076] Step S23: Identify the core feature set We perform multi-scale time series analysis on the time series features in the data and construct derived features, then combine these derived features with... Non-temporal features are merged to form the final enhanced feature matrix.
[0077] 1. Discrete Wavelet Transform (DWT) Decomposition
[0078] For any temporal core feature The discrete wavelet transform (DWT) is used to decompose it into sub-band signals of different frequencies. The DWT is then passed through a set of low-pass filters. and high-pass filter And downsampling operation implementation. Approximation coefficients of layer decomposition (Low frequency) and detail coefficient The formula for (high frequency) calculation is:
[0079] ;
[0080] ;
[0081] in, This is the index of the downsampled coefficients. (This is achieved through...) Layer decomposition, the original signal is represented as:
[0082] ;
[0083] in, It is the first The approximate signal of the layer characterizes the long-term trend of the sequence; These are detailed signals at each level, representing the change patterns at different time scales, from short-term fluctuations to medium-term cycles.
[0084] 2. Construction of temporal features
[0085] Original time series features Replace with the wavelet decomposition obtained from it. Component sequence:
[0086] ;
[0087] These component sequences, as new features with clear physical meaning (trends, periods and noise in different frequency bands), are related to... The merging of non-temporal features ultimately forms an enhanced feature matrix that is richer in information and has stronger representational capabilities. .
[0088] Specifically, step S3 includes the following sub-steps:
[0089] Step S31: Constructing an agricultural product price prediction model using the final enhanced feature matrix obtained after processing in S2. The corresponding agricultural product price series is used as output to train a complete model. Considering the complexity of agricultural product price data, this patent chooses a new BiLSTM model that can comprehensively consider both the past and future of the series.
[0090] 1. BiLSTM cell forward computation
[0091] For the input feature sequence At each time step The LSTM unit updates its state through the following gating mechanism:
[0092] ;
[0093] in, It is the sigmoid activation function. This represents element-wise multiplication. and These are network parameters.
[0094] 2. Two-way information aggregation
[0095] BiLSTM consists of two LSTM layers: a forward layer and a backward layer. These layers process the sequence separately to obtain the forward hidden state sequence. and backward hidden state sequence The final representation of each time step is formed by splicing the two together:
[0096] ;
[0097] 3. Model Training
[0098] The hidden state of the final time step (Or the state after weighted aggregation via an attention mechanism) is passed through a fully connected layer to obtain the price prediction. The model is trained by minimizing the mean squared error (MSE) between the predicted and true values.
[0099] ;
[0100] Step S32:
[0101] 1. System ablation experimental design
[0102] After constructing the preliminary prediction model, in order to verify the effectiveness of the cascaded feature extraction and enhancement method proposed in this patent, a set of systematic and rigorously controlled ablation experiments needs to be designed.
[0103] The specific experimental setup is as follows:
[0104] Control group A (Baseline): Uses the original feature set after S1 preprocessing, without any feature extraction and enhancement processing of this invention.
[0105] Control group B (KNN-MI only): Using the primary feature subset obtained only after step S21 (KNN-MI filtering). .
[0106] Control group C (KNN-MI + XGBoost): using the core feature set obtained after steps S21 and S22. .
[0107] Experimental group D (complete method of this invention): using the enhanced feature matrix obtained after processing through complete steps S21, S22, and S23. .
[0108] 2. Evaluation Indicators and Statistical Analysis
[0109] On the test set, calculate and compare the following performance metrics for each experimental group:
[0110] Root mean square error: ;
[0111] Mean absolute percentage error: ;
[0112] Coefficient of determination: ;
[0113] To further verify the statistical significance of the performance differences, the Diebold-Mariano (DM) test was used. This test compares the prediction error sequences of the two models. To assess whether there are significant differences between them, construct the test statistic:
[0114] ;
[0115] in The loss function (e.g., squared error). Calculation. sample mean and its long-term variance estimation Then the DM statistic is:
[0116] ;
[0117] Under the null hypothesis (that the two models have the same prediction accuracy), the DM statistic asymptotically follows a standard normal distribution. If the p-value is less than the significance level (e.g., 0.05), the null hypothesis is rejected, and it is considered that there is a significant difference in the prediction accuracy of the two models.
[0118] Application concept: Taking soybean price prediction as an example
[0119] S1: Acquisition and preprocessing of multi-source heterogeneous data.
[0120] Soybeans are susceptible to various factors. This patent collects influencing factors from the dimensions of other agricultural product prices, natural environment, macroeconomics, and agricultural futures prices. The data included in each dimension are as follows:
[0121] Agricultural product price dimension: peanut oil, sugarcane, tomatoes, apples, roses, goji berries, live pigs, crucian carp;
[0122] Macroeconomic dimensions: Interest rates, exchange rates, GDP, CRB index, and consumer price index.
[0123] Agricultural futures price dimensions: peanuts, sugar, rapeseed oil, apples, soybeans, and live hogs.
[0124] Natural environmental dimensions: temperature, humidity, daily rainfall, sunshine duration
[0125] All the above data were cleaned, timestamps were aligned, a few missing values were filled using linear interpolation, and Z-score standardization was performed to form a dataset containing... The original feature set of each feature.
[0126] S2: Feature extraction and construction based on a three-level cascaded architecture.
[0127] S21: Primary coarse sieve. Calculates the mutual information between each feature and soybean price. Set the threshold to the average mutual information value, and filter to obtain results containing... Primary subset of features .
[0128] S22: Selected Level 2 Embedded Systems. As a feature, Train an XGBoost model for the labels. Rank the models based on the feature importance (Gain) of their outputs, and select the top-ranked features. The features constitute the core feature set .
[0129] S23: Three-level timing enhancement construction. Identification. The time-series features, such as "mean temperature (laged by 7 days)", are analyzed. A three-level 'db4' wavelet transform is performed on this sequence, yielding one approximate component (low-frequency trend) and three detail components (high-frequency fluctuations). These four new components are then used as new features, and... Non-temporal features are merged to ultimately form a dimension of Enhanced feature matrix .
[0130] S3: Constructing a prediction model and conducting ablation verification experiments.
[0131] Model building: using Train a BiLSTM model as the final predictor.
[0132] Ablation experiment design: The control group A, B, C and experimental group D (in this invention) were strictly set up according to the aforementioned invention description. All groups used the same data partitioning and the same BiLSTM model for training and testing.
[0133] Expected Analysis and Demonstration: On the test set, the following performance ranking is expected to be observed: RMSE(A) > RMSE(B) > RMSE(C) > RMSE(D). This result demonstrates that: (a) any level of feature processing is superior to the original features; (b) secondary selection is superior to primary coarse screening; and (c) the complete framework of this invention exhibits optimal performance. Furthermore, the feature importance of model D in experimental group can be analyzed, and the "temperature trend component" constructed by wavelet is expected to rank highly, explaining its contribution mechanistically.
[0134] S4: Model Evaluation and Application.
[0135] The RMSE, MAE, and other metrics of the experimental group D model on the test set were calculated to evaluate its predicted accuracy. After training, the model can receive new multi-source feature data and output predicted values for future soybean prices for market early warning or decision support.
[0136] It should be noted that the user information (including but not limited to user device information, personal user information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the laws, regulations and standards of relevant countries and regions.
[0137] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.
[0138] 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 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.
[0139] 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. Without further limitation, 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 said element.
[0140] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.
Claims
1. A cascaded feature extraction method for agricultural product price forecasting, characterized in that, Includes the following steps: S1: Acquire multi-source heterogeneous data on agricultural products and preprocess them to form an original feature set; S2: The original feature set is subjected to feature extraction and enhancement using a three-level cascaded architecture to obtain an enhanced feature matrix; S3: Construct a prediction model based on the enhanced feature matrix, and verify the effectiveness of the cascaded architecture through ablation experiments; S4: Use the trained prediction model to predict agricultural product prices.
2. The method according to claim 1, characterized in that, Step S2 employs a three-level cascaded architecture to extract and enhance features from the original feature set, obtaining an enhanced feature matrix, specifically including: S21: Based on the K-nearest neighbor mutual information estimation method, calculate the mutual information value between each feature in the original feature set and the target price index, and obtain a primary feature subset according to a preset threshold or ranking. ; S22: Using the primary feature subset Using the input as input, an XGBoost regression model is trained, and features are extracted based on the importance ranking of split gain. The top-ranked features are selected to form the core feature set. ; S23: For the core feature set Discrete wavelet decomposition is performed on the time series features in the data to extract multi-scale components as derived features, and then combined with... Non-temporal features are merged to form an enhanced feature matrix. .
3. The method according to claim 2, characterized in that, In step S21, the mutual information value is calculated using a nonparametric estimation method based on K-nearest neighbors. The estimation formula is: ； in, For the sample size, For nearest neighbor parameters, It is the Digamma function; The first Each sample in features Univariate space, target variable The corresponding univariate space Number of nearest neighbor samples; Is it the Digamma function in and The function value at that location; It is any one of the influencing factor features in the original feature set for agricultural product price prediction. Target agricultural product prices.
4. The method according to claim 2, characterized in that, In step S22, the feature importance of the XGBoost model is obtained by summing the gain values of the feature at all split nodes in each regression tree, and the calculation formula is as follows: ； in, The total number of trees, Indicates the first Using features in a tree The set of nodes to be split; For the first Tree species characteristics The specific identifier of the split node when executing node splitting; For the first Tree species characteristics At the split node The split gain value corresponding to the split is executed at that point.
5. The method according to claim 2, characterized in that, In step S23, the time series features are decomposed by discrete wavelet transform to obtain the approximate coefficient sequence and the detail coefficient sequence, which respectively represent the long-term trend and the multi-scale fluctuation mode, and these components are added to the feature matrix as new features.
6. The method according to claim 1, characterized in that, In step S3, the prediction model uses a bidirectional long short-term memory network (BiLSTM) and is trained using the mean squared error loss function.
7. The method according to claim 1, characterized in that, The ablation experiment in step S3 includes the following control group: Control group A: Using the original feature set; Control group B: Using a feature subset selected only by KNN-MI ; Control group C: Feature subset selected using KNN-MI and XGBoost ; Experimental group D: Enhanced feature matrix after full three-level processing ; The effectiveness of feature processing at each stage was verified by comparing the prediction performance metrics of each control group on the test set.
8. The method according to claim 7, characterized in that, The performance metrics include root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²), and the statistical significance of the difference in prediction accuracy is verified using the Diebold-Mariano test.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.