Characteristic crop price prediction method and system based on deep learning multi-fusion technology

Abstract

The invention discloses a characteristic crop price prediction method and system based on a deep learning multi-fusion technology, and relates to the technical field of agricultural data analysis, and the method comprises the steps: determining characteristic crop price fluctuation feature information and price influence factors; a multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM is constructed, and the STL-ETO-EMA-PILSTM captures a characteristic crop price complex dynamic state based on the multi-fusion technology and parameter optimization; and according to the characteristic crop price fluctuation characteristic information, the price influence factors and the STL-ETO-EMA-PIL STM, predicting the characteristic crop price. Through feature importance analysis, physical information embedding, a multi-scale attention mechanism and parameter optimization, the characteristic crop price prediction accuracy is improved. Through visual prediction results and analysis, a more reliable decision support tool can be provided for participants, and the participants are helped to better understand the cause and trend of price fluctuation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural data analysis, in particular to a characteristic crop price prediction method and system based on deep learning multi-fusion technology. BACKGROUND

[0002] Agriculture is the foundation of economic and social stability. In recent years, changes in global trade patterns and frequent extreme weather have led to increased price fluctuations for characteristic crops such as garlic, onions, and ginger. These crops have characteristics such as short production cycles, high regional concentration, and limited market capacity. These crops are influenced by factors such as supply and demand imbalance, natural disasters, and speculative behavior, making their pricing mechanism complex. This has brought significant market risks and decision-making challenges to the supply chain. Accurate price prediction can optimize planting decisions, reduce producer risk, guide consumer expectations, and provide a basis for market regulation.

[0003] Characteristic crops exhibit significant price volatility due to seasonal production, regional concentrated supply, and limited storage conditions. Compared to other agricultural products, these characteristics make them more sensitive to multiple factors. Traditional driving factors such as supply and demand balance, climate impact, labor, transportation and storage disruptions, and inflationary pressure exacerbated by interest rate and monetary policy initiatives have pushed up production and storage costs, further amplifying price fluctuations. Digital economic factors such as online public opinion and market expectations have also exacerbated price fluctuations, creating a complex interaction mechanism that makes price volatility uncertainty more complex.

[0004] For the prediction of characteristic crop prices such as garlic, onions, and ginger, existing research mainly uses statistical models such as ARIMA, exponential smoothing, and machine learning methods such as variational mode decomposition based on XGBoost, particle swarm optimization-genetic algorithm, Prophet, GARCH long-short term memory model, normalized long-short term memory model, STL-LSTM-ATT-KAN combination model, and adaptive transformer. While the above statistical models and machine learning methods can achieve characteristic crop price prediction, these methods have strong data dependency, weak cross-regional and crop generalization ability, poor interpretability, limited dynamic feature capture, and slow optimization convergence speed, which affects the accuracy of price prediction. SUMMARY

[0005] To solve the above technical problems, the present application proposes the following technical solutions: In a first aspect, the present application provides a characteristic crop price prediction method based on deep learning multi-fusion technology, comprising: determining characteristic crop price fluctuation feature information and price influencing factors; constructing a multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM; According to the characteristic crop price fluctuation characteristic information and the price influencing factors and the STL-ETO-EMA-PIL STM, the price of the characteristic crop is predicted.

[0006] In a possible implementation, the STL-ETO-EMA-PILSTM comprises a seasonal trend decomposition module STL, a physically constrained long short-term memory network module PILSTM, an efficient multi-scale attention module EMA, and an exponential triangle optimization module ETO. The STL is configured to receive a raw price time series of a characteristic crop as input and split the raw price time series into a trend factor, a seasonal factor, and a residual factor that influence the price. The PILSTM is configured to receive output information of the STL and perform deep-time sequence feature extraction on the received information by using a long short-term memory network LSTM, to output a multi-dimensional time sequence feature vector as input of the EMA module, and to embed a physically constrained mechanism into a structure of the long short-term memory network by combining a physical information neural network PINN on the basis of the deep features extracted by the LSTM, so as to improve the explainability and generalization ability of the model. The EMA is configured to output a weighted price time feature according to the received multi-dimensional time sequence feature vector. The ETO is configured to receive module parameters of the PILSTM and the EMA and perform optimization, and finally feed back to the PILSTM and the EMA respectively, so as to improve the overall prediction performance of the STL-ETO-EMA-PILSTM.

[0007] In a possible implementation, the STL comprises a trend factor extraction submodule, a seasonal factor extraction submodule, and a residual factor extraction submodule. The trend factor extraction submodule calculates a long-term change trend of the price sequence by using a local weighted regression smoothing method, to reflect an overall upward or downward trend, and a formula is as follows: wherein, is the trend factor, is a decomposition function, is a raw price time sequence, is a trend factor smoothing parameter; The seasonal factor extraction submodule identifies annual or quarterly price cycle fluctuations by periodic smoothing operation, and a formula is as follows: wherein, is the seasonal factor, is a seasonal factor smoothing parameter; The residual factor extraction submodule calculates a difference between the raw sequence and the trend factor and the seasonal factor, to obtain a random disturbance part that influences the price.

[0008] In a possible implementation, the PILSTM comprises: a forget gate sub-module for calculating and filtering information in the historical state that needs to be discarded, to avoid irrelevant features from interfering; an input gate sub-module for combining the input features of the current time step and the historical state, to control the introduction and update of new information; an output gate sub-module for determining the state features that are valid information transmitted to the subsequent layer, to realize dynamic modeling of long-term and short-term dependencies; a physical constraint construction sub-module: according to the fluctuation law of agricultural market prices, the price change logic is converted into a differential equation form; a residual calculation sub-module: the automatic differentiation is used to calculate the residual of the predicted output under the physical equation; a constraint loss embedding sub-module: the above residual is embedded as a penalty term into the training loss function of the LSTM, so that the network meets the data fitting accuracy and physical consistency in the parameter update process.

[0009] In a possible implementation, the residual of the predicted output under the physical equation is calculated by using automatic differentiation, comprising: The residual term of the heat conduction equation is used to calculate the automatic differentiation to capture the characteristic crop price fluctuation: ; Further physical capture of the inertia of the characteristic crop price change, the constraint residual loss term of the PINN structure is: Wherein: is the first derivative of the price, indicating the rate of change of the price, is the second derivative of the price, indicating the trend change of the price, is the weight coefficient, controlling the trend sensitivity, is the price of the characteristic crop, is the time variable, is the feature variable of the influencing factor, and N is the sample number, is the predicted price of the th sample.

[0010] In a possible implementation, the EMA outputs the weighted price time features according to the received multi-dimensional time sequence feature vector, comprising: After the EMA receives the multi-dimensional time sequence feature vector from the PILSTM output, it first performs two average pooling operations in the X and Y directions to compress redundant information and extract global statistical features in the spatial dimension, to ensure that different scale time sequence patterns can form effective representation; The pooled result is concatenated with the original features, and cross-channel information fusion and dimensionality compression are achieved through 1×1 convolution, which improves feature interaction efficiency while maintaining channel integrity. In the process of generating attention weights, the Sigmoid function is used to perform non-linear mapping on the convolution output, making the weight distribution smooth and differentiable, and realizing flexible selection of features in key time periods. In the output stage, group normalization is introduced, which divides the channel into multiple subgroups and performs normalization processing separately. Through 1×1 convolution branches and 3×3 convolution branches, local interaction information within the channel and multi-scale patterns of local regions are captured respectively. Finally, through matrix dot product, the attention maps output by the two branches achieve pixel-level interaction modeling, and are then merged into a channel-weighted map.

[0011] In one possible implementation, group normalization is introduced at the output stage, dividing the channel into multiple subgroups for standardization. 1×1 and 3×3 convolutional branches are used to capture local interaction information within the channel and multi-scale patterns in local regions, respectively, as shown in the formula: in, This represents the attention map of the output of a 1×1 convolutional branch. This represents the attention map of the output of a 3×3 convolution branch.

[0012] In one possible implementation, the ETO receives and optimizes the module parameters of the PILSTM and the EMA, and finally feeds them back to the PILSTM and the EMA respectively, to improve the overall prediction performance of the STL-ETO-EMA-PILSTM, including: After receiving the module parameters of the PILSTM and the EMA, iterative optimization is performed, and a fitness value is calculated in each optimization iteration. The fitness value measures the quality of an individual in the solution space and serves as the core basis for subsequent search and updates. Subsequently, a triangular perturbation search mechanism guided by triangular geometry is used, and a position update formula is constructed by combining exponential and trigonometric functions. A phased switching mechanism is adopted to dynamically transition between exploration and development, thereby enhancing the stability of the search and the global optimization capability. An exponentially decaying cosine triangle perturbation fine-tuning mechanism is introduced. By superimposing small-amplitude random perturbations on the cosine fluctuations, the parameters achieve gradually shrinking convergence as they approach the optimal solution. The trigonometric functions still participate in the perturbation, but their amplitudes are gradually compressed, thus achieving fine-grained parameter search and eventual convergence. After each parameter optimization update, the fitness value of the candidate solution is recalculated and compared with the current best solution. If the fitness value is better, the global best solution is updated as the optimization parameter and fed back.

[0013] In one possible implementation, the optimization process is divided into two phases, including: During the exploration phase, a large-scale search of the solution space is conducted to prevent getting trapped in local optima. The formula is as follows: During the development phase, a fine-grained search is performed around the current optimal solution region to improve convergence accuracy. The formula is as follows: in It is the first The parameter in the first... The result of the second optimization. This is a constant used to control the search intensity and fluctuation range. To introduce periodic fluctuations, This represents the current optimal solution. and These are coefficients used to control the convergence speed and the perturbation frequency. Indicates the maximum number of optimization attempts. This indicates an increase in periodic local disturbances. The representative exponential decrease factor controls the magnitude of the disturbance to decrease over time.

[0014] In one possible implementation, the ETO uses a stage scheduling factor to control the transition between two stages of the optimization process, expressed as: Where t is the current number of optimizations, and T is the maximum number of optimizations. It reflects the current timeline of the exploration.

[0015] Secondly, embodiments of this application provide a special crop price prediction system based on deep learning multi-fusion technology, characterized in that it includes: The determination module is used to identify the price fluctuation characteristics and price influencing factors of specialty crops; The model building module is used to build a multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM, which captures the complex dynamics of specialty crop prices based on multi-fusion technology and parameter optimization. The price prediction module is used to predict the price of the specialty crops based on the price fluctuation characteristics and price influencing factors of the specialty crops and the STL-ETO-EMA-PIL STM.

[0016] In this embodiment, a prediction model is developed based on the acquired characteristic information of specialty crop price fluctuations and price influencing factors, combined with a multi-fusion technology capable of capturing the complex dynamics of specialty crop prices. Through feature importance analysis, physical information embedding, multi-scale attention mechanisms, and parameter optimization, the accuracy of specialty crop price predictions is improved. By providing intuitive prediction results and analysis, it also offers participants more reliable decision support tools, helping them better understand the causes and trends of price fluctuations. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for predicting the price of specialty crops based on deep learning multi-fusion technology, provided for an embodiment of this application; Figure 2 A schematic diagram of the STL-ETO-EMA-PILSTM structure provided in the embodiments of this application; Figure 3 A comparative schematic diagram of the garlic price prediction models provided in the embodiments of this application; Figure 4 This is a comparative diagram of the onion price prediction models provided in the embodiments of this application; Figure 5 This is a comparative diagram of the ginger price prediction models provided in the embodiments of this application; Figure 6 A schematic diagram of a specialty crop price prediction system based on deep learning multi-fusion technology provided in this application embodiment. Detailed Implementation The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0018] See Figure 1 The specialty crop price prediction method based on deep learning multi-fusion technology provided in this embodiment includes: S101, determine the price fluctuation characteristics and price influencing factors of specialty crops.

[0019] The price fluctuations of specialty crops exhibit complex characteristics, with their long-term trends significantly influenced by factors such as supply and demand, and seasonality. Planting area and yield are particularly important. To construct an accurate prediction model, this embodiment employs a moving average method to smooth the price data, effectively suppressing short-term fluctuations. The original dataset exhibits clear periodic and monthly fluctuations. After data processing, the trends become clearer and more stable. Preliminary analysis indicates that specialty crop prices are influenced by a combination of long-term trends, seasonal factors, and short-term unforeseen events. To support the theoretical construction of the prediction model, further research is needed on the interaction mechanisms among these multiple factors.

[0020] The price fluctuations of specialty agricultural crops are influenced by a variety of factors, including traditional elements (supply and demand, climate conditions) as well as emerging variables (production costs, macroeconomic indicators, market dynamics). Market speculation and major events not only directly affect prices but also indirectly impact them by increasing agricultural input costs. Notably, online public opinion has been proven to have a price fluctuation transmission effect. To improve forecast accuracy, this embodiment systematically reviews key influencing factors, aiming to build a more accurate price forecasting model. Detailed analysis is shown in Table 1.

[0021] Table 1. Main Influencing Factors on the Price of Specialty Crops This embodiment employs the Pearson correlation coefficient method to construct a variable correlation heatmap. By quantifying the strength of linear correlations among various factors, it deeply analyzes the influencing factors and their interactions on the price fluctuations of specialty crops. This accurately identifies the key drivers affecting price fluctuations and effectively avoids multicollinearity issues. These statistical results provide a scientific basis and variable selection foundation for building robust price forecasting models, thereby improving the model's interpretability and predictive accuracy.

[0022] After determining the price fluctuation characteristics and price influencing factors of specialty crops, the original price time series of specialty crops is generated with time on the horizontal axis and price on the vertical axis, taking into account the above influencing factors.

[0023] S102, construct a multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM.

[0024] See Figure 2 The STL-ETO-EMA-PILSTM constructed in this embodiment includes: a seasonal trend decomposition module STL, a physically constrained long short-term memory network module PILSTM, an efficient multi-scale attention module EMA, and an exponential triangular optimization module ETO.

[0025] The STL (Standardized Time Series) receives the original price time series of specialty crops as input and decomposes it into trend factors, seasonal factors, and residual factors affecting prices. The PILSTM (Physical Information Network for Short-Term Memory) receives the output information from the STL and uses a Long Short-Term Memory (LSTM) network to perform deep temporal feature extraction on the received information, outputting a multi-dimensional temporal feature vector as input to the EMA (Economic Model for Short-Term Memory) module. Furthermore, based on the deep features extracted by the LSTM, a Physical Information Neural Network (PINN) is combined to embed a physical constraint mechanism into the structure of the LSTM network, thereby improving the model's interpretability and generalization ability. The EMA outputs a weighted price time feature based on the received multi-dimensional temporal feature vector. The ETO (Electronic Tolerance) receives the module parameters of the PILSTM and EMA, optimizes them, and finally feeds them back to the PILSTM and EMA respectively, thereby improving the overall predictive performance of the STL-ETO-EMA-PILSTM model.

[0026] The STL decomposition technique is employed to decompose the price series into three parts: trend factors reflecting long-term factors, seasonal factors representing cyclical patterns, and residual factors capturing sudden events. This effectively solves the inherent coupling problem in multi-scale features. Secondly, an innovative EMA multi-scale attention mechanism is proposed, which uses three sets of convolutional kernels to model short-term features (weather, public trends), medium-term features (policy, inventory), and long-term features (production capacity, consumption). This design is expected to significantly improve the model's responsiveness to external factors. The PINN physical constraint mechanism is introduced. This mechanism embeds the supply and demand balance equation into the loss function, thus forming a prediction process that conforms to market equilibrium principles. Model optimization is achieved using an improved ETO algorithm, which effectively promotes the dynamic equilibrium of the parameter search process.

[0027] In this embodiment, the STL includes a trend factor extraction submodule, a seasonal factor extraction submodule, and a residual factor extraction submodule. The PILSTM includes: a forget gate submodule, an input gate submodule, an output gate submodule, a physical constraint construction submodule, a residual calculation submodule, and a constraint loss embedding submodule.

[0028] The price fluctuations of specialty crops are influenced by a combination of factors, including historical price volatility, climate conditions, holiday effects, and seasonal supply and demand, exhibiting significant time dependence and nonlinear characteristics. Traditional linear models often fail to adequately capture the long-term or short-term dynamic patterns when modeling such complex nonlinear fluctuations.

[0029] Therefore, this embodiment proposes a Long Short-Term Memory (LSTM) modeling method. As an improved version of Recurrent Neural Networks (RNNs), LSTM regulates the retention and forgetting of information through a gating mechanism, enabling it to capture long-term dependencies in time-series data, thereby overcoming the limitations of traditional linear model prediction accuracy. The LSTM network structure can effectively cope with the complexity of price fluctuations in specialty crops, providing a practical solution for improving price prediction accuracy.

[0030] LSTM has been proven to achieve adaptive storage of historical information, dynamic fusion of current information, and efficient extraction of hidden states through its unique triple-gate structure. In commodity price series modeling, this network has been shown to effectively identify long-term trends and short-term fluctuations, thereby significantly improving the representation efficiency of time series features.

[0031] LSTM possesses the ability to retain key information over a long period and achieves rapid state updates through a weight adjustment mechanism. The LSTM forget gate... Control over old memories The retention is achieved through the input gate. Controlling new information The addition of, as in the formula: In research on specialty crop price forecasting, this mechanism enables the model to comprehensively consider both long-term trends and short-term fluctuations. The model can accurately update its state at each time step based on current input data and historical memory information, as shown in the formula: Output gate It controls which parts of the memory can be exposed. It is the final effective feature representation, which can be regarded as the model's comprehensive understanding of historical price behavior. The output of the LSTM unit at the current time step represents the extraction and encoding of deep temporal features of the historical input sequence.

[0032] To enhance the model's generalization ability, a PINN structure based on physical information constraints derived from differential equations is introduced. PILSTM, on the other hand, incorporates the PINN physical constraint mechanism into the basic LSTM structure. PINN exhibits unique advantages in handling problems involving physical phenomena, particularly in capturing the inherent laws of complex physical systems and making accurate predictions of unknown data. This network automatically satisfies physical conservation laws, ensuring that the model output matches physical reality, thereby improving the model's reliability and accuracy. PINN demonstrates good adaptability to high-dimensional problems and outperforms traditional numerical methods in terms of computational efficiency.

[0033] While LSTM has certain advantages in capturing long-term dependencies in time series, its performance remains unsatisfactory when handling multi-scale data. This is particularly true when distinguishing between short-term and long-term information, making it difficult to accurately capture price fluctuations. To address this issue, this embodiment proposes an efficient multi-scale attention mechanism aimed at optimizing model performance.

[0034] Multi-scale attention mechanisms are a crucial technique in deep learning, adaptively capturing information at different scales and assigning weights according to importance. Mathematically, they are based on complex function transformations and matrix operations, generating multi-scale feature representations through multi-level feature extraction, providing rich information for subsequent processing. Computationally, they leverage efficient algorithms for rapid feature selection and weight calculation, reducing computational costs and memory consumption, and improving model performance and generalization ability.

[0035] The flexibility of multi-scale attention mechanisms allows them to be adjusted according to task and data characteristics, adapting to different application scenarios. In price prediction, it focuses on key time points and trend changes in price fluctuations. It can be combined with other deep learning techniques, such as convolutional neural networks, to enhance the model's expressive power.

[0036] To enhance the LSTM's response to recent price dynamics, this embodiment introduces an efficient multi-scale attention mechanism module, EMA, which weights the hidden states at different time steps, making the model more attentive to recent significant fluctuations. Specialty crop price fluctuations are influenced by various factors, exhibiting weekly, monthly, and seasonal trends as well as sudden disturbances. EMA's multi-branch parallel processing capability can effectively smooth unstructured disturbances, making it suitable for high-resolution feature representation and fine-grained modeling tasks, particularly well-suited for unstructured, highly volatile agricultural data. This invention embeds the EMA module into the LSTM structure for specialty crop price prediction, improving performance without significantly increasing the training burden.

[0037] The performance of prediction models is significantly affected by multiple parameters, which may have complex nonlinear dependencies. Therefore, introducing an optimization algorithm with high efficiency and strong global optimization capabilities is crucial in this embodiment. The ETO algorithm is an emerging swarm intelligence innovation optimization algorithm inspired by the triangular geometric transformation concept and exponential weighting strategy. Its core idea is to simulate the "triangular geometric structure between an individual point and two guiding points" and combine it with an exponential weighting method to dynamically update the search space, thereby achieving a balance between global exploration and local development capabilities for the optimal solution.

[0038] S103, based on the price fluctuation characteristics and price influencing factors of the specialty crops and STL-ETO-EMA-PILSTM, the price of the specialty crops is predicted.

[0039] After the model is built, the original price time series of the specialty crops identified in S101 are input into STL. The trend factor extraction submodule of STL uses a locally weighted regression smoothing method to calculate the long-term trend of the price series, reflecting the overall upward or downward trend. The formula is as follows: ,in: As a trend factor, For decomposition function, The original price time series, The smoothing parameter is used for trend factors. The seasonality factor extraction submodule identifies annual or quarterly price cycle fluctuations through periodic smoothing operations, using the following formula: ,in: Due to seasonal factors, The seasonality factor smoothing parameter is used; the residual factor extraction submodule calculates the difference between the original sequence and the trend factor and seasonal factor to obtain the random disturbance part affecting the price.

[0040] STL successfully separates seasonal and trend factors from the data, significantly enhancing the accuracy and stability of the STL-ETO-EMA-PILSTM model in forecasting. This technique enables subsequent LSTMs to more accurately capture dependencies in time-series data. The PILSTM includes a forgetting gate module, which calculates and filters information from historical states to avoid interference from irrelevant features. An input gate module combines the input features of the current time step with historical states to control the introduction and updating of new information. An output gate module determines the state features that are passed to subsequent layers as valid information, enabling dynamic modeling of long-term and short-term dependencies. A physical constraint construction submodule transforms the price fluctuation logic into differential equations based on agricultural market price fluctuation patterns. A residual calculation submodule uses automatic differentiation to calculate the residuals of the predicted output under the physical equations. A constraint loss embedding submodule embeds the residuals as penalty terms into the LSTM's training loss function, ensuring that the network simultaneously satisfies data fitting accuracy and physical consistency during parameter updates.

[0041] Furthermore, the residual calculation submodule utilizes automatic differentiation to predict the residuals of the output under the physical equations, including: calculating automatic differentiation using the residual terms of the heat conduction equation to capture price fluctuations of specialty crops. .

[0042] To further capture the inertia of price changes in specialty crops, the constrained residual loss term of the PINN structure is: in: Let be the first derivative of the price, and represent the rate of change of the price. Let be the second derivative of price, representing the trend change in price. These are weighting coefficients that control trend sensitivity. For the price of specialty crops, For time variables, Here, N represents the characteristic variables of the influencing factors. For the first Predicted prices for each sample.

[0043] By incorporating differential equations as prior physical knowledge into the neural network training process, the PINN model not only fits historical data but also reveals underlying temporal evolution patterns, significantly enhancing its generalization ability. In research on price forecasting for minor agricultural commodities, PINN aims to construct a differential control equation that approximates dynamic price changes and incorporates the residuals of this equation as a penalty term into the training process to ensure that the network learning follows the principle of physical consistency. By integrating data-driven methods with physical modeling, the PINN structure significantly improves the interpretability of minor agricultural commodity price forecasting models, ensuring that the model's learning process follows specific economic logic and evolutionary patterns.

[0044] After receiving the multidimensional temporal feature vector from the PILSTM output, the EMA first performs two average pooling operations in the X and Y directions to compress redundant information and extract global statistical features in the spatial dimension, ensuring that effective representations can be formed for temporal patterns at different scales. The pooled result is concatenated with the original features, and cross-channel information fusion and dimensionality compression are achieved through 1×1 convolution, improving feature interaction efficiency while maintaining channel integrity. In the process of generating attention weights, the Sigmoid function is used to perform non-linear mapping on the convolution output, making the weight distribution smooth and differentiable, and enabling flexible selection of features in key time periods.

[0045] To further improve the model's stability under high-dimensional features, group normalization is introduced in the output stage. Channels are divided into multiple subgroups for standardization, mitigating variance fluctuations caused by mini-batch training and ensuring consistency in feature distribution across different batches. 1×1 and 3×3 convolutional branches are used to capture local interaction information within channels and multi-scale patterns in local regions, respectively, enhancing the model's responsiveness to local patterns and thus better capturing long-term dependencies in the data. The formula is: ,in: This represents the attention map of the output of a 1×1 convolutional branch. This represents the attention map output by the 3×3 convolution branch. Finally, through matrix dot product, the attention maps output by the two branches achieve pixel-level interaction modeling, and are then merged into a channel-weighted map.

[0046] As explained in S102, the performance of the prediction model is significantly affected by multiple parameters, and these parameters may have complex nonlinear dependencies. Therefore, parameter optimization is crucial to the accuracy of model predictions.

[0047] In this embodiment, ETO receives the module parameters from PILSTM and EMA and performs iterative optimization, calculating a fitness value in each iteration. This fitness value measures the merit of an individual in the solution space and serves as the core basis for subsequent search updates. A triangular perturbation search mechanism guided by triangular geometry is then used, combining exponential and trigonometric functions to construct a position update formula. A phased switching mechanism dynamically transitions between exploration and development, enhancing search stability and global optimization capabilities. An exponentially decaying cosine triangular perturbation fine-tuning mechanism is introduced. By superimposing small-amplitude random perturbations on cosine fluctuations, the parameters achieve gradually decreasing convergence as they approach the optimal solution. Trigonometric functions still participate in the perturbation, but their amplitudes are gradually compressed, achieving fine-grained parameter search and eventual convergence. After each parameter optimization update, the fitness value of a candidate solution is recalculated and compared with the current optimal solution. If its fitness is better, the global optimal solution is updated as the optimization parameter for feedback.

[0048] In this implementation, the ETO optimization process is divided into two stages, including: During the exploration phase, a large-scale search of the solution space is conducted to prevent getting trapped in local optima. The formula is as follows: During the development phase, a fine-grained search is performed around the current optimal solution region to improve convergence accuracy. The formula is as follows: in It is the first The parameter in the first... The result of the second optimization. This is a constant used to control the search intensity and fluctuation range. To introduce periodic fluctuations, This represents the current optimal solution. and These are coefficients used to control the convergence speed and the perturbation frequency. Indicates the maximum number of optimization attempts. This indicates an increase in periodic local disturbances. The representative exponential decrease factor controls the magnitude of the disturbance to decrease over time.

[0049] In this embodiment, ETO uses a stage scheduling factor to control the transition between two stages of the optimization process, expressed as: Where t is the current number of optimizations, and T is the maximum number of optimizations. It reflects the current timeline of the exploration.

[0050] ETO (Extended Search Tolerance) is used for fine-grained control of the search process, dynamically adjusting global and local search weights. The bidirectional perturbation mechanism of the triangular structure design enhances the algorithm's ability to explore the solution space, increases diversity, and avoids local optima. This algorithm is suitable for nonlinear, multimodal, and complex function optimization, efficiently exploring the search space, reducing resource consumption, shortening solution time, and ensuring the discovery of the global optimum. The exponential decay strategy aligns with agricultural market patterns, adapting to seasonal price fluctuations and ensuring long-term model performance.

[0051] To verify the above prediction method, the following experiments were also conducted in this embodiment: The experiment was conducted on a SUSE ST-SN260 server equipped with an AMD EPYC 9754 processor (128 cores, 2.25GHz), an NVIDIA RTX 4060 graphics card (8GB VRAM), 384GB of Samsung DDR5 memory, a 2TB SSD, and a 16TB HDD. The system ran Ubuntu 20.04 LTS, with Python 3.8.10 installed and core libraries such as PyTorch 1.9.0, NumPy 1.21.0, and Pandas 1.3.0 configured.

[0052] This study obtained agricultural product price data from 2014 to 2024 from the National Key Agricultural Product Market Information Platform of the Ministry of Agriculture and Rural Affairs of China, focusing on the daily wholesale prices of three specialty crops (garlic, onion, and ginger). Data collection was conducted in collaboration with authoritative institutions to ensure the reliability of the research support. Publicly available data was obtained through web scraping technology; static pages were scraped using Python's Requests and BeautifulSoup libraries, while dynamic content requiring login was scraped using Selenium in conjunction with XPath.

[0053] During data preprocessing, the garlic, scallion, and ginger datasets were carefully cleaned and organized to ensure high data quality. The specific steps are as follows: 1. Then, the dataset is refined to eliminate duplicates and resolve outliers.

[0054] 2. Use the average imputation method of the corresponding category to fill in the missing values ​​in the dataset to ensure data integrity and consistency.

[0055] 3. To ensure a seamless analysis process, the data undergoes optimization. Before prediction, all input features are standardized—a process proven to effectively improve the stability and convergence efficiency of neural networks. After prediction, the system performs inverse standardization, restoring the predicted values ​​and actual prices to their original magnitudes, facilitating accurate comparison with real data.

[0056] 4. By applying STL time series decomposition technology, specialty crop price data can be broken down into three components: trend, seasonality, and residuals. This method transforms complex price fluctuation patterns into more easily analyzable elements, enabling subsequent models to accurately capture long-term trends and cyclical changes in the data.

[0057] 5. Constructing a time-series sliding window is crucial, and the original time-series data needs to be converted into the format required for supervised learning. Subsequently, each consecutive historical feature is processed as an input sequence, ultimately forming a three-dimensional input structure (number of samples, time step, number of features) that meets the requirements of an LSTM network.

[0058] The implementation of these preprocessing steps has been shown to enhance the reliability and applicability of the resulting datasets.

[0059] This dataset contains 4,018 records from January 1, 2014 to December 31, 2024, and has been carefully divided into an 80% training set and a 20% test set based on time series characteristics. This division maximizes the use of training data to reduce the risk of overfitting while ensuring a sufficient test set containing unseen data to improve the model's generalization ability.

[0060] To evaluate the model's predictive performance on specialty crop prices, we use four core metrics: Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). MAE, calculated as the average absolute difference between predicted and actual values ​​(as shown in the formula below), directly reflects the average deviation between the model's predicted values ​​and the actual prices.

[0061] Among them, it means Actual value Indicates the predicted value. It refers to the sample size.

[0062] Mean squared error (MSE) is the average of the squared errors, emphasizing larger deviations, as shown below: The root mean square error (RMSE) is calculated as the square root of the mean squared difference between the observed and actual values, as shown below: The coefficient of determination (R²) quantifies the proportion of the variance of the dependent variable that the model explains, reflecting its quality of fit, as shown below: The value of is between 0 and 1, which directly reflects the model's fit quality. The closer the value is to 1, the stronger the explanatory power of the model.

[0063] A comprehensive evaluation of time-series forecasting models for specialty crop prices requires the application of multi-scale attention mechanisms and physical information neural networks, along with the validation and evaluation methods described in this paper. Overcoming existing limitations and optimizing model architecture and parameter settings are crucial for improving forecast accuracy and practical value. This research approach provides solid scientific support for stabilizing specialty crop markets and guiding policy decisions.

[0064] To ensure optimal performance of all model components, the Exponential Triangular Optimization (ETO) algorithm was used for precise parameter tuning in this experiment. While optimizing hyperparameters using ETO, other parameters such as the time window were manually adjusted. Each prediction used a randomly generated parameter configuration, and detailed records were kept for each set of prediction parameters. See Table 2 for specific parameter configurations.

[0065] Table 2 shows the parameter selection for the model. To accurately evaluate the predictive performance of the model, four evaluation indicators—mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²)—were used to quantitatively analyze the price prediction results for garlic, onions, and ginger. The analysis results are shown in Table 3.

[0066] Table 3. Prediction Results of Specialty Crops The smaller the values ​​of MAE, MSE, and RMSE, the higher the prediction accuracy, while the closer the R² value is to 1, the better the model fit.

[0067] As shown in Table 3, the empirical data demonstrates the model's excellent predictive ability. Garlic performed best, with the lowest mean error indices (MAE=0.0853, MSE=0.0145, RMSE=0.1192), followed by scallions (MAE=0.0581, MSE=0.0062, RMSE=0.0792). Although ginger's error was larger due to drastic price fluctuations (MAE=0.1409, MSE=0.0363, RMSE=0.1905), it remained within acceptable limits. All models had R² values ​​close to 1 (garlic: 0.9975, scallions: 0.9978, ginger: 0.9882), fully validating its explanatory power for price fluctuation patterns. It is particularly important to emphasize that a higher R² value indicates optimal model fit, while a decrease in error indices directly corresponds to improved predictive accuracy.

[0068] To comprehensively evaluate the performance of the models in prediction tasks, this study conducted comparative experiments using six pricing prediction models: Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Lasso Regression, Random Forest (RF), Gradient Boosting, and LightGBM. The specific experimental results are as follows.

[0069] Table 4 Comparison of Garlic Price Forecast and Experimental Results Table 5 Comparison of Experimental Results for Onion Price Prediction Table 6 Comparison of Experimental Results for Ginger Price Prediction Table 4 shows the performance comparison of various models in garlic price prediction. STL-ETO-EMA-PILSTM performs excellently in evaluation metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²), fully demonstrating its effectiveness. This model exhibits significant advantages among mainstream machine learning models in this field: its MAE value of 0.0853 is significantly lower than Support Vector Machine (SVM)'s 2.2236 and XGBoost's 0.1461, strongly validating its reliability in price prediction. Furthermore, the R² value reaches 0.9975, indicating a high degree of consistency between the predicted price and the actual price trend, highlighting the model's interpretability and stability. The core advantage of this model lies in its accurate simulation of seasonal fluctuations and the market inertia characteristics of garlic prices: STL decomposition technology extracts price trends and cyclical components; the EMA module captures price-sensitive fluctuations within specific time periods; and PINN-LSTM enhances its responsiveness to sudden market factors by introducing physical constraints, ensuring robustness in dynamic environments.

[0070] As shown in Table 5, the STL-ETO-EMA-PILSTM model demonstrates superior performance in predicting onion prices, with a mean absolute error (MAE) of 0.0581 and root mean square errors (RMSE) of 0.0062 and 0.0792, respectively. Compared to advanced models such as LightGBM and XGBoost, this method shows a significant performance improvement, with an R² value of 0.9978, fully demonstrating its ability to accurately capture onion price trends. Unlike the relatively stable price of garlic, onion prices are more volatile and significantly influenced by regional climate patterns and sales channels. Traditional models often face overfitting or underfitting problems when dealing with such highly volatile data. This model integrates market dynamics from different time dimensions through the multi-scale attention mechanism of the EMA module and combines it with continuous modeling techniques under physical constraints to ensure that the prediction results are both stable and conform to physical laws.

[0071] As shown in Table 6, the STL-ETO-EMA-PILSTM model demonstrates superior performance in ginger price forecasting. Specifically, its mean absolute error (MAE) is the lowest at 0.1409, significantly outperforming XGBoost (0.3623) and LightGBM (0.2706). Its mean squared error (MSE) is 0.0350, and its root mean square error (RMSE) is 0.1871, representing a substantial reduction in prediction error compared to traditional models. Furthermore, the coefficient of determination (R²) reaches 0.9967, further validating the model's ability to accurately reconstruct price trend patterns. Ginger prices exhibit significant seasonality, with supply information lags during the planting season and market supply and demand adjustments also exhibiting a certain degree of lag. These factors place higher demands on the accuracy of price forecasting models. To address this challenge, this model employs Time Series Decomposition (STL) for effective seasonal analysis, utilizes Long Short-Term Memory (LSTM) networks to model temporal dependencies, and leverages an efficient multi-scale attention mechanism (EMA) to highlight abrupt changes. Meanwhile, physical information neural network (PINN) constraints are integrated into the modeling process to regulate the persistence of price fluctuation patterns, thereby achieving accurate modeling of complex nonlinear dynamics.

[0072] See Figures 3 to 5 These are comparison charts of garlic, scallion, and ginger price prediction models. The horizontal axis represents time, and the vertical axis represents price. The True curve represents the actual price, while the other curves represent prices not predicted by the models. The remaining curves represent the prediction curves of STL-ETO-EMA-PILSTM, Lasso, Support Vector Machine (SVM), XGBoost, RF, Gradient Boosting, and LightGBM, respectively.

[0073] Depend on Figures 3 to 5 The comparative results show that STL-ETO-EMA-PILSTM demonstrates superior performance in predicting the prices of garlic, onions, and ginger. This model has two significant advantages: first, it accurately captures real-time price movements, especially excelling at price fluctuation inflection points; second, it effectively reflects the characteristics of price volatility, maintaining high prediction accuracy even during periods of sharp fluctuations. Compared to other models such as Lasso, SVM, XGBoost, Random Forest, Gradient Boosting, and LightGBM, this model provides more realistic predictions in most time windows, fully demonstrating its ability to handle complex time series data. Furthermore, the model exhibits excellent stability and robustness during the prediction process, with smooth prediction curves and no significant fluctuations or outliers.

[0074] Price forecasts for garlic, onions, and ginger show that STL-ETO-EMA-PILSTM not only effectively predicts price trends but also demonstrates high stability and robustness. These findings further validate the model's broad applicability and efficiency as an advanced forecasting tool. Experiments demonstrate that STL-ETO-EMA-PILSTM, as a robust time series forecasting tool, provides reliable support for agricultural product price forecasting.

[0075] This experiment also evaluated the effectiveness of the model framework in predicting specialty crop prices through a series of ablation experiments, comparing the EMA-PINN-LSTM model with the traditional LSTM model. This model innovatively combines an efficient multi-scale attention module with a physically based LSTM architecture. This series of experiments aims to verify whether relying solely on the EMA module and the PINN-LSTM structure can still maintain the model's predictive performance when global trend modeling capabilities are lacking.

[0076] This study also evaluated the performance of the traditional LSTM model in complex price prediction tasks, particularly in scenarios lacking multi-scale feature extraction and trend analysis capabilities. By comparing the prediction results of the two models, this study accurately analyzed the contribution of each component to the overall performance and provided empirical evidence for model optimization. Detailed data are shown in Tables 7-9.

[0077] Table 7 Results of the garlic price prediction ablation experiment Table 8 Results of the ablation experiment for predicting onion prices Table 9 Results of the ablation experiment for predicting ginger prices Table 7 shows the experimental results of the garlic price prediction ablation experiment. By removing model components, the contribution of each module to the prediction performance can be clearly distinguished. STL-ETO-EMA-PILSTM performs excellently on all indicators, surpassing not only the standard EMA-PINN-LSTM model but also the pure LSTM model. Existing research shows that STL technology has been proven to accurately capture the long-term trend and cyclical fluctuations of garlic prices, effectively mitigating the bias caused by LSTM when processing non-stationary data. The EMA module effectively captures short-term price fluctuations, thereby improving the model's sensitivity. The PINN-LSTM architecture, by introducing physical constraints, enhances the model's ability to cope with abnormal fluctuations, ensuring prediction stability. Although garlic prices exhibit regular characteristics, seasonal changes and extreme situations can still cause instability in traditional models. The ETO-optimized STL-EMA-PINN-LSTM model improves prediction accuracy while providing high-precision support for the decision-making process.

[0078] Table 8 shows the experimental results of the study on onion price forecasting. The STL-EMA-PINN-LSTM model performed best, with a mean absolute error (MAE) of 0.0581; followed by the EMA-PINN-LSTM model with an error of 0.0624; the pure LSTM model had an error as high as 2.2502. The complete model achieved an R² value of 0.9978, which not only surpassed other models but also verified the performance improvement brought about by modular integration. Onion prices fluctuate due to various factors, and the STL and EMA modules play a crucial role in ensuring stable and reliable forecast results. The accuracy of the model directly affects the effectiveness of supply chain planning; this advanced model effectively improves supply chain efficiency and risk management capabilities through high-precision forecasting.

[0079] Table 9 presents the regression analysis results for ginger price forecasting. The mean absolute error (MAE) of the STL-ETO-EMA-PILSTM model was 0.1409. After removing the trend decomposition module, the MAE increased to 0.1601, while the MAE of the pure Long Short-Term Memory (LSTM) model significantly climbed to 1.6960. Notably, the coefficient of determination (R²) of the complete model reached 0.9967, far exceeding the 0.4184 R² of the independent LSTM model. This finding fully demonstrates the superior predictive power of the integrated architecture.

[0080] Corresponding to the above embodiment of the specialty crop price prediction method based on deep learning multi-fusion technology, this application also provides an embodiment of a specialty crop price prediction system based on deep learning multi-fusion technology.

[0081] See Figure 6 The specialty crop price prediction system 20 based on deep learning multi-fusion technology in this embodiment includes: Module 201 is used to determine the price fluctuation characteristics and price influencing factors of specialty crops.

[0082] The model building module 202 is used to build a multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM, which captures the complex dynamics of specialty crop prices based on multi-fusion technology and parameter optimization.

[0083] The price prediction module 203 is used to predict the price of the specialty crop based on the price fluctuation characteristics and price influencing factors of the specialty crop and the STL-ETO-EMA-PIL STM.

[0084] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0085] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for predicting the price of specialty crops based on deep learning multi-fusion technology, characterized in that, include: Identify the characteristics of price fluctuations and factors influencing prices of specialty crops; A multi-fusion technology combination prediction model STL-ETO-EMA-PILSTM is constructed. The STL-ETO-EMA-PILSTM captures the complex dynamics of specialty crop prices based on multi-fusion technology and parameter optimization. Based on the price fluctuation characteristics and price influencing factors of the specialty crops, and using STL-ETO-EMA-PIL STM, the prices of specialty crops are predicted.

2. The method for predicting specialty crop prices based on deep learning multi-fusion technology according to claim 1, characterized in that, The STL-ETO-EMA-PILSTM includes: a seasonal trend decomposition module STL, a physically constrained long short-term memory network module PILSTM, an efficient multi-scale attention module EMA, and an exponential triangular optimization module ETO. The STL is used to receive the original price time series of specialty crops as input and decompose it into trend factors, seasonal factors and residual factors that affect prices; The PILSTM is used to receive the output information of the STL and uses the Long Short-Term Memory (LSTM) network to perform deep temporal feature extraction on the received information, outputting a multi-dimensional temporal feature vector as the input of the EMA module. Furthermore, based on the deep features extracted by the LSTM, the Physical Information Neural Network (PINN) is combined to embed the physical constraint mechanism into the structure of the LSTM network, thereby improving the interpretability and generalization ability of the model. The EMA outputs a weighted price-time feature based on the received multidimensional time-series feature vector; The ETO is used to receive and optimize the module parameters of the PILSTM and the EMA, and finally feed them back to the PILSTM and the EMA respectively, in order to improve the overall prediction performance of the STL-ETO-EMA-PILSTM.

3. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 2, characterized in that, The STL includes a trend factor extraction submodule, a seasonal factor extraction submodule, and a residual factor extraction submodule; The trend factor extraction submodule uses a local weighted regression smoothing method to calculate the long-term trend of the price series, reflecting the overall upward or downward trend. The formula is as follows: ,in: As a trend factor, For decomposition function, The original price time series, For smoothing parameters of trend factors; The seasonality factor extraction submodule identifies annual or quarterly price cycle fluctuations through periodic smoothing operations, using the following formula: ,in: Due to seasonal factors, For smoothing parameters of seasonal factors; The residual factor extraction submodule calculates the difference between the original sequence and the trend and seasonal factors to obtain the random disturbance component affecting the price.

4. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 2, characterized in that, The PILSTM includes: The forgetting module is used to calculate and filter information that needs to be discarded in historical states, avoiding interference from irrelevant features; The input gate module is used to combine the input features of the current time step with the historical state to control the introduction and updating of new information; The output gate submodule is used to determine the state characteristics that are passed to subsequent layers as valid information, thereby enabling dynamic modeling of long-term and short-term dependencies. Physical constraint construction submodule: Based on the price fluctuation patterns in the agricultural market, the price change logic is transformed into differential equation form; The residual calculation submodule uses automatic differentiation to predict the residuals of the output under the physical equations. Constraint loss embedding submodule: The above residuals are embedded as a penalty term into the training loss function of LSTM, so that the network can simultaneously satisfy data fitting accuracy and physical consistency during parameter update.

5. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 4, characterized in that, The method of using automatic differentiation to calculate the residual of the predicted output under the physical equations includes: Automatic differential calculations using the residual terms of the heat conduction equation to capture price fluctuations in specialty crops: ; To further capture the inertia of price changes in specialty crops, the constrained residual loss term of the PINN structure is: in: Let be the first derivative of the price, and represent the rate of change of the price. Let be the second derivative of price, representing the trend change in price. These are weighting coefficients that control trend sensitivity. For the price of specialty crops, For time variables, Here, N represents the characteristic variables of the influencing factors. For the first Predicted prices for each sample.

6. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 2, characterized in that, The EMA outputs a weighted price-time feature based on the received multidimensional time-series feature vector, including: After receiving the multidimensional temporal feature vector output from the PILSTM, the EMA first performs two average pooling operations in the X and Y directions to compress redundant information and extract global statistical features in the spatial dimension, ensuring that effective representations can be formed for temporal patterns at different scales. The pooled result is concatenated with the original features, and cross-channel information fusion and dimensionality compression are achieved through 1×1 convolution, which improves feature interaction efficiency while maintaining channel integrity. In the process of generating attention weights, the Sigmoid function is used to perform non-linear mapping on the convolution output, making the weight distribution smooth and differentiable, and realizing flexible selection of features in key time periods. In the output stage, group normalization is introduced, which divides the channel into multiple subgroups and performs normalization processing separately. Through 1×1 convolution branches and 3×3 convolution branches, local interaction information within the channel and multi-scale patterns of local regions are captured respectively. Finally, through matrix dot product, the attention maps output by the two branches achieve pixel-level interaction modeling, and are then merged into a channel-weighted map.

7. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 6, characterized in that, The output stage introduces group normalization, dividing the channels into multiple subgroups for standardization. 1×1 and 3×3 convolutional branches are used to capture local interaction information within the channels and multi-scale patterns in local regions, respectively. The formula is as follows: in, This represents the attention map of the output of a 1×1 convolutional branch. This represents the attention map of the output of a 3×3 convolution branch.

8. The method for predicting specialty crop prices based on deep learning multi-fusion technology according to claim 2, characterized in that, The ETO receives and optimizes the module parameters of the PILSTM and the EMA, and finally feeds them back to the PILSTM and the EMA respectively, in order to improve the overall prediction performance of the STL-ETO-EMA-PILSTM, including: After receiving the module parameters of the PILSTM and the EMA, iterative optimization is performed, and a fitness value is calculated in each optimization iteration. The fitness value measures the quality of an individual in the solution space and serves as the core basis for subsequent search and updates. Subsequently, a triangular perturbation search mechanism guided by triangular geometry is used, and a position update formula is constructed by combining exponential and trigonometric functions. A phased switching mechanism is adopted to dynamically transition between exploration and development, thereby enhancing the stability of the search and the global optimization capability. An exponentially decaying cosine triangle perturbation fine-tuning mechanism is introduced. By superimposing small-amplitude random perturbations on the cosine fluctuations, the parameters achieve gradually shrinking convergence as they approach the optimal solution. The trigonometric functions still participate in the perturbation, but their amplitudes are gradually compressed, thus achieving fine-grained parameter search and eventual convergence. After each parameter optimization update, the fitness value of the candidate solution is recalculated and compared with the current best solution. If the fitness value is better, the global best solution is updated as the optimization parameter and fed back.

9. The method for predicting the price of specialty crops based on deep learning multi-fusion technology according to claim 8, characterized in that, The optimization process consists of two phases, including: During the exploration phase, a large-scale search of the solution space is conducted to prevent getting trapped in local optima. The formula is as follows: During the development phase, a fine-grained search is performed around the current optimal solution region to improve convergence accuracy. The formula is as follows: in It is the first The parameter in the first... The result of the second optimization. This is a constant used to control the search intensity and fluctuation range. To introduce periodic fluctuations, This represents the current optimal solution. and These are coefficients used to control the convergence speed and the perturbation frequency. Indicates the maximum number of optimization attempts. This indicates an increase in periodic local disturbances. The representative exponential decrease factor controls the magnitude of the disturbance to decrease over time.

10. The method for predicting the price of specialty crops based on deep learning multi-fusion technology as described in requirement 9, characterized in that: The ETO uses a phase scheduling factor to control the transition between two phases of the optimization process, expressed as: Where t is the current number of optimizations, and T is the maximum number of optimizations. It reflects the current timeline of the exploration.

Images