Technology and AI method for generating digital and intelligent quotation through agricultural supply chain production place field market transaction

Abstract

The invention provides an agricultural supply chain production place field market transaction generation digital-intelligent quotation technology and an AI method, and relates to the technical field of agricultural supply chain transaction data resource utilization research and development, AI decision making and transaction price artificial intelligence prediction. The method is a digital intelligent technology innovation aiming at an agricultural supply chain production place field end product marketing transaction activity. The method mainly comprises the following steps: a transaction vector metadata data processing [3: 3: 3: + / -1] mixed structure of an artificial intelligence model of a digital intelligence technology: vector metadata interconnection of an input layer, hidden layers I and II and an output layer; the AI system physical assembly comprises an information acquisition module, a data processing and representing module, an intelligent calculation module, a feedback and optimization module, a multi-stage transaction supporting module and a modular interface, and replacement and upgrading of the calculation module, the data processing module and the transaction supporting module are achieved. And the artificial neuron performs reverse propagation training and weight adjustment on the model by using a BP algorithm so as to perform continuous optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence transaction quotation prediction in agricultural supply chain, in particular to a kind of agricultural supply chain production site field market transaction generates digital quotation technology and AI method. BACKGROUND

[0002] With the rapid development of fresh agricultural products industry, the transaction management of production site field market plays an increasingly important role in improving farmers' income, optimizing market resource allocation and supporting government policy implementation. Modern field market needs to collect supply chain vector transaction data in real time, accurately predict quotations and quickly respond to market fluctuations to cope with the growing transaction demand and complex market environment. However, traditional field market transaction management technology has significant limitations in data collection, quotation setting and market response, which is difficult to meet the needs of diversified scenarios such as peak transaction period, off-season low period and sudden events (such as weather disasters).

[0003] Although the existing technology plays a certain role in field market transaction management, there are still the following key problems to be solved:

[0004] Insufficient quotation prediction accuracy: traditional transaction record system relies on single data source (such as manual record or simple statistics), which can only provide historical price or rough trend, and cannot integrate multi-source data (such as production cost, market demand, government subsidy) for dynamic quotation prediction, resulting in inability to accurately identify price fluctuations caused by supply and demand changes or regional differences. At the same time, lack of confidence interval support, unable to give early warning of price risk 15-30 minutes in advance, resulting in chaotic peak transaction, farmers difficult to optimize income.

[0005] Low intelligence of quotation strategy: existing strategy formulation relies on personal experience or simple data analysis, and cannot develop customized quotation based on multi-source data (such as seasonal demand, yield change). For example, off-season quotation is mostly fixed downward (e.g., 10% discount), which lacks pertinence, resulting in low transaction success rate (usually less than 60%), which is difficult to effectively improve farmers' income or market activity.

[0006] Low efficiency of resource allocation: traditional technology lacks multi-source vector metadata integration (such as yield, price, environmental data), which is difficult to dynamically optimize transaction resource allocation. For example, transportation vehicles or storage space are based on fixed plan, which cannot be adjusted accurately according to real-time transaction volume (such as daily transaction volume exceeding 1000 kg) or market changes, often leading to idle resources in off-season or insufficient transportation in peak period.

[0007] Market fluctuation response speed is slow (days): manual intervention in market price fluctuation processing depends on subjective judgment, and the response time is long (days), and there is a lack of systematic and rapid adjustment mechanism. For example, when heavy rain causes supply interruption, manual telephone notification takes a long time, and it is difficult to complete adjustment within a short time (minutes), increasing the risk of transaction failure and economic loss.

[0008] Therefore, it is necessary to apply agricultural supply chain vector metadata intelligent technology and AI method innovation to solve the above problems. SUMMARY

[0009] TECHNICAL SCHEME

[0010] To achieve the above purpose, the following intelligent agricultural supply chain technical scheme is used: an agricultural supply chain origin field market transaction generation intelligent pricing technology and AI method, an artificial intelligence model includes an input layer, a neuron system and an output layer, the neuron system includes a first hidden layer, a second hidden layer, a historical / similar data set and a confidence interval calculation unit, wherein:

[0011] The input layer is used to collect feature vector metadata related to field market transactions, including price, yield, region, user identity and other information, and receive multi-source original vector metadata, including production information of marketable agricultural products, market demand information, price information and cost reference information provided by the government of the production area;

[0012] The first hidden layer is used for feature extraction and selection from the multi-source vector metadata data of the input layer, extracting feature parameters related to production cost, marketable price, market supply and demand, season, regional demand and government regulation measures, performing feature selection to remove irrelevant features, and normalizing the extracted features;

[0013] The second hidden layer is used for feature construction and enhancement based on the output of the first hidden layer, including combined features, derived features, weight adjustment and dynamic update, and associated calculation with historical data and similar samples to generate a higher fitting degree of pricing prediction results;

[0014] The confidence interval calculation unit is used to calculate the confidence interval of the transaction price based on the output result of the second hidden layer, and the confidence interval calculation includes but is not limited to confidence estimation methods based on statistical distribution, sampling reconstruction and Bayesian analysis;

[0015] The output layer is used to output transaction pricing results and their confidence intervals based on forward propagation algorithm, activation function and back propagation algorithm, including confidence intervals of different transaction modes, and the pricing results are suitable for different transaction modes.

[0016] Preferably, the input layer can collect the feature data through mobile terminals, web pages, IoT sensors, and blockchain data interfaces.

[0017] Preferably, the feature extraction process of the first hidden layer includes feature selection and optimization steps based on correlation analysis, dimensionality reduction analysis, and self-learning network structure, so as to improve the effectiveness of feature representation and model generalization ability.

[0018] Preferably, the standardization process of the first hidden layer includes calculating the mean, variance, and interval range based on the sample distribution characteristics to achieve numerical scale uniformity among different feature dimensions.

[0019] Preferably, the second hidden layer dynamically adjusts the parameters by constructing time series features and fluctuation features, combined with an error feedback mechanism, in order to optimize the fitting accuracy of the price prediction model.

[0020] Preferably, when generating a confidence interval, the confidence interval calculation unit calculates the confidence boundary based on the sample variance, sample size, and error range, and can automatically select a suitable statistical model according to the data characteristics.

[0021] Preferably, the transaction quotes and confidence intervals output by the output layer are applicable to various market models, including counterparty trading, auction trading, competitive bidding, and online matching.

[0022] Preferably, the method includes the following steps:

[0023] Data collection phase: Collect characteristic data of fresh agricultural product market transactions, including price, output, region and user identity, and receive raw metadata, including production metadata of agricultural products entering the market, market demand and price information metadata, and production cost information metadata of the producing area government;

[0024] Feature processing stage: The metadata is processed by a neural system, which includes a first hidden layer, a second hidden layer, a historical / similar dataset, and a confidence interval calculation unit, wherein:

[0025] The first hidden layer uses correlation analysis, linear discriminant analysis and unsupervised autoencoder algorithm for feature extraction and selection. The autoencoder includes an encoder, a decoder and a loss function, and uses Z-Score normalization, sample mean calculation, sample standard deviation calculation and degree of freedom calculation for feature normalization.

[0026] Model optimization phase: The second hidden layer constructs and enhances features based on the output of the first hidden layer, including calculating the moving average and volatility of prices, and updating the weights using mean squared error, backpropagation error and gradient descent, and combining them with the historical / similar dataset;

[0027] Confidence interval calculation stage: The confidence interval calculation unit calculates confidence intervals based on the output of the second hidden layer, including t-distribution confidence intervals, proportional confidence intervals, Bootstrap method and Bayesian confidence intervals, wherein the t-distribution confidence interval involves the t-distribution critical value tα / 2,df, error range and confidence interval boundary;

[0028] Results output stage: Based on the forward propagation algorithm, activation function and back propagation algorithm, output the trading quote results and their confidence intervals, including confidence intervals for different trading modes.

[0029] Beneficial effects

[0030] This invention provides a digital pricing technology and AI method for generating market transactions at the agricultural production site in the supply chain. It has the following beneficial effects:

[0031] 1. In this invention, the system collects multi-source characteristic supply chain vector metadata (such as price, output, region, and user identity) and raw metadata (such as production costs, market demand, and government subsidy information) in real time from the field market through the input layer. It then applies multi-layer processing of an artificial neural network system to achieve dynamic price prediction. Compared to the limitations of traditional transaction record systems that rely on single historical data, this system can accurately identify price fluctuations caused by changes in supply and demand and regional differences. Through a confidence interval calculation unit, it provides a price range with a 95% confidence level and supports price risk warnings 15-30 minutes in advance, thereby effectively reducing transaction chaos during peak periods and helping farmers optimize their income.

[0032] 2. In this invention, the system utilizes artificial intelligence technology in the first and second hidden layers to extract, construct, and enhance data functions, and formulates customized pricing strategies based on multi-source vector metadata (such as seasonal demand and production changes). Compared to the current situation where experience-driven fixed discounts (such as a 10% reduction) result in low transaction success rates (below 60%), the system significantly improves transaction success rates (up to 80% or more) by dynamically adjusting weights and derived features (such as moving averages and volatility), effectively increasing farmers' income and market activity, with target growth rates of 10% and 5%, respectively.

[0033] 3. In this invention, the system integrates multi-source vector metadata such as output, price, and environment collected from the input layer, and combines it with the dynamic update mechanism of the neural network system to achieve intelligent allocation of trading resources. Compared with traditional technologies that rely on fixed plans, resulting in idle resources during the off-season or insufficient transportation during peak periods, the system can dynamically adjust transportation vehicles or storage space based on real-time trading volume (e.g., daily trading volume exceeding 1000 kg), optimizing resource utilization, reducing waste, and improving service capacity during peak periods.

[0034] 4. In this invention, the system achieves rapid response to market fluctuations by combining forward and backward propagation algorithms at the output layer. Compared to the current situation where manual intervention is time-consuming (on the order of hours) and lacks a systematic adjustment mechanism, the system can complete price adjustments and notifications within minutes. For example, when heavy rain causes supply disruptions, it automatically updates prices and pushes them to traders via the APP / Web platform, significantly reducing the risk of transaction failures and alleviating operational pressure. Attached Figure Description

[0035] Fig. 1 This is a topological diagram of the present invention;

[0036] Fig. 2 This is a schematic diagram of the physical component architecture of the system of the present invention;

[0037] Fig. 3 This is a technical system framework diagram of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Specific Implementation Example 1: As shown in the example Figs. 1-3 As shown,

[0040] A digital pricing technology and AI method for generating farm-to-farm market prices in the agricultural supply chain, corresponding to the farm-to-farm market price declaration stage, includes the following technical solutions: input, hidden layer, and output layer.

[0041] The input layer is used to collect feature data related to the transaction of fresh agricultural products in the market, including information such as price, output, region, and user identity, and to receive multi-source raw metadata, including production information, market demand information, price information and cost reference information provided by the local government of the producing area for agricultural products entering the market.

[0042] The first hidden layer is used to extract and select features from the multi-source data of the input layer. It extracts feature parameters related to production costs, market supply and demand, seasons, regional demand and government regulation measures, performs feature selection to remove irrelevant features, and normalizes the extracted features.

[0043] The second hidden layer is used to construct and enhance features based on the output of the first hidden layer, including combining features, deriving features, weight adjustment and dynamic updates, and performing correlation calculations with historical data and similar samples to generate price prediction results with higher fit.

[0044] The confidence interval calculation unit is used to calculate the confidence interval of the transaction quote based on the output of the second hidden layer. The confidence interval calculation includes, but is not limited to, confidence estimation methods based on statistical distribution, sampling reconstruction and Bayesian analysis.

[0045] The output layer is used to output trading quotes and their confidence intervals based on the forward propagation algorithm, activation function, and back propagation algorithm. The confidence intervals include those for different trading modes, and the quotes are applicable to different trading modes.

[0046] The input layer can collect feature data through mobile terminals, web pages, IoT sensors, and blockchain data interfaces.

[0047] The feature extraction process of the first hidden layer includes feature selection and optimization steps based on correlation analysis, dimensionality reduction analysis, and self-learning network structure, in order to improve the effectiveness of feature representation and the model's generalization ability.

[0048] The standardization process of the first hidden layer includes calculating the mean, variance, and interval range based on the sample distribution characteristics, so as to achieve numerical scale uniformity among different feature dimensions.

[0049] The second hidden layer constructs time series features and volatility features, and dynamically adjusts the parameters by combining an error feedback mechanism to optimize the fitting accuracy of the price prediction model.

[0050] When generating confidence intervals, the confidence interval calculation unit calculates the confidence boundaries based on the sample variance, sample size, and error range, and can automatically select an appropriate statistical model based on the data characteristics.

[0051] The transaction quotes and confidence intervals output by the output layer are applicable to various market models, including counterparty trading, auction trading, competitive bidding, and online matching.

[0052] Execution steps:

[0053] Sp1: Feature-based metadata data collection stage: Collect feature data of fresh agricultural product market transactions, including price, output, region and user identity, and receive raw metadata, including production metadata of agricultural products entering the market, market demand and price information metadata, and production cost information metadata of the producing area government.

[0054] Sp2: Feature Vector Metadata Processing Stage: Metadata is processed through a neuron system, which includes a first hidden layer, a second hidden layer, historical / similar datasets, and confidence interval calculation units, where:

[0055] The first hidden layer uses correlation analysis, linear discriminant analysis and unsupervised autoencoder algorithm for feature extraction and selection. The autoencoder includes an encoder, a decoder and a loss function, and Z-Score normalization, sample mean calculation, sample standard deviation calculation and degree of freedom calculation are used for feature normalization.

[0056] Model optimization phase: The second hidden layer constructs and enhances features based on the output of the first hidden layer, including calculating the moving average and volatility of prices, and updating the weights using mean squared error, backpropagation error and gradient descent, and combining them with historical / similar datasets;

[0057] Confidence interval calculation stage: The confidence interval calculation unit calculates the confidence interval based on the output of the second hidden layer, including the t-distribution confidence interval, the proportional confidence interval, the Bootstrap method, and the Bayesian confidence interval. The t-distribution confidence interval involves the t-distribution critical value tα / 2, df, the error range, and the confidence interval boundary.

[0058] Sp3: Feature Vector Metadata Output Stage: Based on the forward propagation algorithm, activation function, and backpropagation algorithm, output the transaction quote results and their confidence intervals, including confidence intervals for different transaction modes.

[0059] Specific Implementation Example 2: As shown in the example Figs. 1-3 As shown, the following is the content of the AI ​​method in Example 1:

[0060] The main content of the input layer is:

[0061] Mobile terminals, web pages, IoT sensors, and blockchain data interfaces collect characteristic data on fresh agricultural product market transactions, including prices, output, regions, and user identities.

[0062] The input layer receives raw metadata, namely, metadata of raw information data on the production, market, and government departments of agricultural products entering the market, including raw information metadata on production costs Q1; raw information metadata on demand and price of agricultural products entering the market Q2; and raw information metadata on production costs of the producing area government Q3.

[0063] The main contents of the first and second hidden layer neuron systems are as follows:

[0064] Feature extraction: Extract production-related cost data, market supply and demand data related to seasons and regions, and government adjustment and subsidy control measures data from the input layer metadata. These will directly affect production costs and subsequent market quotations.

[0065] Feature selection: Based on correlation analysis and feature uniqueness, select features that are highly correlated with transaction quotes and remove features that are not highly correlated or irrelevant.

[0066] Feature normalization: Normalize the extracted features, such as through standardization of metrics and indices, so that features with different dimensions and value ranges have the same scale.

[0067] Main contents of the output layer neuron system:

[0068] Output results: These outputs are the optimal results obtained through forward propagation, combined with activation functions and backpropagation (BP) optimization. The output layer outputs the final transaction quote results, including the confidence intervals of the transaction quotes, and obtains the confidence intervals for different transaction modes (counterpartie transactions, auction education).

[0069] Specific Implementation Example 3: As shown in the example Figs. 1-3 As shown, the physical components of the AI ​​method in the above embodiments are configured as follows: Fig. 2 As shown:

[0070] The hardware components of the input layer include: HTTP services based on Nginx / Express / Apache, website and mobile applications; data storage MySQL; distributed computing Spark; cloud services AWS; and open-source computing technologies: Pandas / NumPy / Reguests / Flask.

[0071] The hardware components of the first hidden layer include: computing framework: TensorFlow / PyTorch; data storage and retrieval: Redis / HDFS; neural network parameter storage: Cassandra; distributed computing: Spark; and open-source computing technologies: TensorFlow / PyTorch / Keras.

[0072] The hardware components of the second hidden layer include: computing framework: TensorFlow / PyTorch; data storage and retrieval: Redis / HDFS; neural network parameter storage: Cassandra; distributed computing: Spark; and open-source computing technologies: TensorFlow / PyTorch / Keras.

[0073] The hardware components of the output layer include: display panel: Grafana / Tableau, etc.; real-time feedback: Apache Kafka / AWS Kinesis; data storage: MySQL / AWs Cloud; open-source computing technology: Matplotlib / Plotly. Specific Implementation Example 4:

[0075] The following is a complete description of the AI ​​methodology's workflow, combined with a real-world application scenario (e.g., a farmer selling fresh vegetables, such as tomatoes, at a market in a southern production area). It further illustrates that the entire process of digital intelligence technology is automated, typically completed within seconds to minutes, depending on the amount of data and computational complexity. The system can be deployed on a cloud server, and users can interact with it through an app or web interface.

[0076] Sp1: Vector metadata acquisition phase (input layer):

[0077] Operating steps:

[0078] The system collects characteristic data of fresh agricultural product market transactions through multiple channels, including price (historical or current quotes), output (current batch output), region (geographical information of the place of origin, such as latitude / longitude or administrative division), and user identity (seller / buyer role, such as farmer ID or purchaser authentication).

[0079] It also receives raw metadata, including production metadata of agricultural products entering the market (Q1, such as planting costs, harvest time, quality indicators, such as organic certification), market demand and price information metadata (Q2, such as current market supply and demand, historical transaction records), and production cost information metadata of the producing area government (Q3, such as subsidy policies, transportation costs, tax adjustments).

[0080] Data collection tools: mobile APP (farmers scan QR codes with their mobile phones to upload yield data), web system (buyers log in to view market dynamics), sensor equipment (IoT devices automatically monitor field yield, humidity and other environmental data).

[0081] Practical application example:

[0082] Farmers use an app to take photos and upload information about tomato yield (e.g., 500 kg) and quality indicators at the field market. The system automatically retrieves regional weather data from sensors (e.g., in South China, affected by the rainy season). Simultaneously, the system pulls subsidy information from government databases (e.g., a subsidy of 0.5 yuan per kg) and demand data from market APIs (e.g., high demand in surrounding cities). This data serves as the raw input, forming a multi-dimensional vector (e.g., [yield: 500, historical price: 3.5, region: South China, ...]).

[0083] If the data is incomplete, the system will prompt the user to supplement it or use the default value to ensure the completeness of the input.

[0084] Sp2: Vector metadata processing stage (Neural system: first hidden layer and second hidden layer)

[0085] First hidden layer: Feature extraction, selection, and normalization.

[0086] Operating steps:

[0087] Extract key features from the metadata of the input layer: features related to production costs (e.g., planting costs), market supply and demand (e.g., demand vs. supply), season (e.g., rainy season causes production fluctuations), regional demand (e.g., South China prefers fresh vegetables), and government regulation measures (e.g., subsidies reduce costs).

[0088] Feature selection is performed: correlation analysis is used to remove irrelevant features (e.g., excluding irrelevant weather data), linear discriminant analysis (LDA) is used to reduce dimensionality to highlight class distinctions (e.g., different types of agricultural products), and unsupervised autoencoders learn latent representations (encoders compress data, decoders reconstruct data, and loss functions minimize reconstruction errors).

[0089] Feature normalization: Z-Score standardization is used (calculate the sample mean, standard deviation and degrees of freedom df=n-1) to make the feature scale consistent (e.g., convert output from tons to standard scores).

[0090] Practical application example:

[0091] After receiving tomato data, the system extracts features such as "high seasonal demand (abundant supply during the rainy season)" and "government subsidies reducing costs" in the first hidden layer. LDA analysis distinguishes between "high-demand areas" and "low-demand areas," and the autoencoder extracts latent patterns (such as cost-yield correlations) from unlabeled production data. Z-Score standardizes the price (range of 3-5 yuan) to the [-1,1] range to avoid large yield values ​​dominating. After processing, a refined feature set (e.g., 10 dimensions reduced to 5 dimensions) is output and passed to the second hidden layer.

[0092] Second hidden layer: Feature construction, enhancement, and preliminary confidence calculation:

[0093] Operating steps:

[0094] Feature construction is based on the output of the first hidden layer: combined features (e.g., creating a "supply and demand index" by multiplying output by seasonal demand), and derived features (e.g., calculating a moving average of prices: the average price over the past 7 days; volatility: standard deviation to measure price stability).

[0095] Feature enhancement: Weight adjustment (highlighting important features, such as the impact of subsidies on pricing), dynamic updates (adjusting weights in real time as new data comes in, using mean squared error (MSE) to assess bias, backpropagating error to calculate gradients, and updating weights and biases using gradient descent).

[0096] Combined with historical / similar datasets: The system calls the stored historical data (e.g., transaction records of similar agricultural products in the past year) or similar datasets (e.g., vegetable data from neighboring regions) to initially calculate the basis of the confidence interval (e.g., sample mean and standard deviation).

[0097] Practical application example:

[0098] For tomatoes, the second hidden layer constructs derived features such as "7-day moving average price = 3.8 yuan" and "volatility = 0.2 (low volatility)". If market data is updated (e.g., a sudden increase in demand), the system uses MSE to calculate the current model error (e.g., predicted price vs. historical price deviation of 0.1), backpropagates the error to the hidden layer, and updates the weights using gradient descent (learning rate η = 0.01), strengthening the weight of the "demand index". Combined with historical datasets (e.g., similar vegetable prices from last month), an enhanced feature set is output, ready for confidence interval calculation.

[0099] Sp3: Confidence Interval Calculation Unit: Calculates Quotation Uncertainty.

[0100] Operating steps:

[0101] Based on the output of the second hidden layer, select an appropriate method to calculate the confidence interval: t-distribution confidence interval (small sample, calculate the critical value tα / 2, df, error range ME = t × s / √n, boundary = mean ± ME); proportion confidence interval (for proportion data, such as demand percentage); Bootstrap method (non-parametric resampling, generating 1000+ subsamples to calculate percentile intervals); Bayesian confidence interval (combining prior distribution, such as historical subsidy prior, to calculate the posterior interval).

[0102] Choose the appropriate distribution based on the data: use the t-distribution for small samples, the Bootstrap distribution for non-normal samples, and the Bayesian distribution if prior knowledge is required.

[0103] Practical application example:

[0104] The system calculates the 95% confidence interval (α=0.05) for tomato prices using a t-distribution (df=sample size-1=49), ME=0.3, with a boundary value of [3.5, 4.1] yuan / kg. If the data is skewed, the Bootstrap method is switched to resample historical data to obtain a more robust interval [3.4, 4.2]. The output to the output layer supports uncertainty quantification.

[0105] Sp3: Model optimization phase (throughout the neural system):

[0106] Operating steps:

[0107] The algorithm uses forward propagation (which computes activation functions layer by layer from the input layer, such as ReLU which introduces nonlinearity) and backpropagation (the BP algorithm optimizes weights and minimizes loss).

[0108] The system is continuously trained: initially, it is trained offline using historical data, and then fine-tuned online during runtime (e.g., new transaction data is used to update the model).

[0109] The significance level α = 1 - (right-left) is used for statistical testing to ensure the reliability of features and intervals.

[0110] Practical application example:

[0111] When processing tomato data, forward propagation calculates an initial price quote, while backpropagation adjusts parameters based on the MSE error. If the prediction bias is large (e.g., ignoring subsidies), the BP is iteratively updated until convergence. The entire optimization process runs in the cloud, requiring no user intervention.

[0112] Sp4: Result Output Stage (Output Layer):

[0113] Operating steps:

[0114] Based on the optimized model, the transaction quote results and their confidence intervals are output, including confidence intervals for different transaction modes (e.g., counterparty transaction: one-to-one quote range; auction transaction: starting price range).

[0115] Output format: Numerical price (e.g., 3.8 yuan / kg) + range (e.g., [3.5, 4.1]) + explanation (e.g., "Based on high current demand, 95% confidence").

[0116] Practical application example:

[0117] Farmers view the output on the app: "Suggested price for tomatoes: 3.8 yuan / kg, 95% confidence interval [3.5, 4.1], applicable to counterparty transactions (refer to the lower limit during negotiation) or auctions (starting price 3.5)." Buyers can see the same result through the web interface and match accordingly. The system records transaction feedback for future dynamic updates.

[0118] Important considerations for the overall practical application of the system:

[0119] Deployment environment: The system runs on a cloud platform (such as Alibaba Cloud), the user end is an APP / Web, and the sensors are integrated with IoT. Data privacy protection: User identity is encrypted, and metadata is anonymized.

[0120] Performance and scalability: Processes single queries in <1 second, and large datasets in <1 minute. Scalable to multiple agricultural products (e.g., adding a fruit module) by updating historical datasets.

[0121] Benefits: In field markets, it reduces the risk of subjective pricing, improves transaction efficiency (e.g., a 20% increase in transaction rate), and supports government regulation (such as subsidies that have a real-time impact on pricing).

[0122] Specific Implementation Example 5: As shown in the example Figs. 1-3 As shown, the following are examples of specific techniques and methods used by farmers in the production areas:

[0123] Background for use:

[0124] A farmer named Li grows and sells fresh tomatoes at a field market in a southern production area. Currently, it's the rainy season, market demand is high, and the government is providing subsidies. Li wants to quickly obtain a smart price quote through the system and then trade with buyers at the field market or participate in an online auction platform. The following is the complete application process of the system in this specific use case.

[0125] Use cases of digital technologies and AI methodologies:

[0126] Sp1: Data Acquisition Phase (Input Layer)

[0127] Xiao Li opened the mobile app on his phone and logged into his farmer account (user ID: seller ID 12345). He clicked the "Start Quoting" button, and the system prompted him to enter the current tomato yield (manually enter 500 kg) and quality grade (select "premium"). At the same time, the app automatically connected to the sensor equipment installed in the field, collected real-time environmental data (such as humidity 80%, temperature 25°C) and recorded the production area (South China).

[0128] The system retrieves subsidy information (0.5 yuan per kilogram) from the government database and demand data (high demand in surrounding cities, with an estimated daily transaction volume of 2,000 kilograms) through a web-based backend. This data forms the raw metadata, including production data (e.g., planting cost of 2 yuan / kg), market demand data (e.g., current market price of 3.5-4 yuan / kg), and government cost data (e.g., transportation subsidy of 0.2 yuan / kg).

[0129] If Xiao Li does not enter the production output, the app will display a "Data Missing" message, suggesting that she select a default value from the history (e.g., the last input was 400 kg) or manually add the data. After Xiao Li confirms, the data collection is complete.

[0130] Sp2: Data Processing Stage (Neural System)

[0131] The first hidden layer: The system automatically processes the collected data and extracts key information related to pricing. For example, it identifies "high demand during the rainy season" and "government subsidies reducing costs" as important features. The system removes irrelevant data (such as irrelevant weather indicators) and then standardizes the features (such as price and output) to make data of different units (such as kilograms and yuan) on the same scale, facilitating subsequent analysis. The processed data is then simplified into a few core features and passed to the next layer.

[0132] The second hidden layer: The system constructs new features based on refined data, such as calculating the moving average price over the past 7 days (approximately 3.8 yuan / kg) and price volatility (low volatility 0.2). It dynamically adjusts feature weights, highlighting the impact of subsidies and demand, and combines historical data (e.g., tomato trading records from last year's rainy season) to initially estimate the price range. The system continuously updates the weights to adapt to real-time market changes.

[0133] Confidence Interval Calculation Unit: The system analyzes the current data sample (assuming 50 recent transactions) and uses the t-distribution method to calculate the 95% confidence interval. Combined with historical data, the system further validates the results to ensure the robustness of the price range.

[0134] SP3: Model Optimization Phase

[0135] The system runs forward and backward propagation algorithms in the background to optimize the neural network parameters. The initial model was trained based on historical data from the past year, and is currently being fine-tuned online using Xiao Li's real-time data. If market anomalies occur (such as a sudden 50% price increase), the system will trigger an alert, prompting the administrator to adjust the model parameters. The optimization process is completed in the cloud, requiring no user intervention from Xiao Li.

[0136] Sp4: Result Output Stage (Output Layer)

[0137] The system generated the final quote, which was displayed on Xiao Li's app's "Quote Details" page: the suggested price was 3.8 yuan / kg, with a 95% confidence interval of [3.5, 4.1] yuan / kg. The page also provided an explanation: "Based on current high demand and subsidies, 95% confidence." Xiao Li saw two suggested transaction models:

[0138] Counterparty transactions: Prices can be negotiated with buyers, with a reference minimum of 3.5 yuan / kg as the starting point.

[0139] Auction transactions: A starting price of 3.5 yuan / kg can be set on the online platform to attract bids.

[0140] Xiao Li selected a counterparty for the transaction, clicked the "Share Quote" button, and sent the result to Mr. Zhang, a nearby buyer. Mr. Zhang received the quote via the web interface and agreed to the transaction at 3.7 yuan / kg. After the transaction was completed, the app popped up a "Feedback Survey," and Xiao Li gave it a 4-star rating (indicating satisfaction). The system recorded the feedback for future optimization.

[0141] Summary of the system's user experience among farmers:

[0142] 1. Time efficiency: From the moment Xiao Li clicks "Start Quotation" to the display of the results, it only takes about 3 seconds (single query response time).

[0143] 2. Convenience: Xiao Li does not need to manually calculate or refer to market conditions. The system automatically integrates data from multiple sources, reducing subjective errors.

[0144] 3. Production income: Compared with the traditional price (e.g., 3 yuan / kg), the system's suggested price of 3.8 yuan / kg increases Xiao Li's income by about 900 yuan (500 kg × 0.8 yuan).

[0145] 4. System expansion: The next time Xiao Li sells other agricultural products (such as cucumbers), he only needs to update the yield and quality data, and the system will automatically adapt.

[0146] It should be noted that, in this document, relational terms such as "including" and "first" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital pricing technology and AI method for generating agricultural supply chain field-to-market transactions. Its features include: The model includes an input layer, hidden layer, and output layer of an artificial neuron model, as well as a hybrid structure of the artificial neuron [3:3:3:±1]: This includes vector metadata for the model structure's price generation, historical / similar datasets, and confidence interval calculation units, wherein: The input layer is used to collect feature data related to agricultural product market transactions, including price, yield, region and user identity information, and to receive multi-source raw metadata, including production information, market demand information, price information and cost reference information provided by the local government of the producing area for agricultural products entering the market. The first hidden layer is used to extract and select features from the input multi-source data, analyze feature parameters related to production costs, market supply and demand, seasonal factors, regional demand and government regulation measures, remove irrelevant features and perform data standardization processing; The second hidden layer is used to construct and enhance features based on the output of the first hidden layer, including combining, deriving and dynamically adjusting market features, and performing correlation calculations with historical data and similar samples to generate price prediction results with higher fit. The confidence interval calculation unit is used to calculate the confidence interval of the transaction quote based on the output of the second hidden layer. The confidence interval calculation includes, but is not limited to, confidence estimation methods based on statistical distribution, sampling reconstruction and Bayesian analysis. The output layer is used to output the transaction quote results and their confidence intervals through the model forward propagation and backward optimization process. The quote results are applicable to different transaction modes.

2. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to claim 1, characterized in that: The input layer can collect the feature data through mobile terminals, web pages, IoT sensors, and blockchain data interfaces.

3. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to claim 1, characterized in that: The feature extraction process of the first hidden layer includes feature selection and optimization steps based on correlation analysis, dimensionality reduction analysis, and self-learning network structure, in order to improve the effectiveness of feature representation and the model's generalization ability.

4. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to claim 1, characterized in that: The standardization process of the first hidden layer includes calculating the mean, variance, and interval range based on the sample distribution characteristics, so as to achieve numerical scale unification among different feature dimensions.

5. The digital pricing technology and AI method for generating agricultural supply chain field market transactions according to claim 1, characterized in that: The second hidden layer constructs time series features and fluctuation features, and dynamically adjusts the parameters by combining an error feedback mechanism to optimize the fitting accuracy of the price prediction model.

6. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to claim 1, characterized in that: When generating confidence intervals, the confidence interval calculation unit calculates the confidence boundaries based on the sample variance, sample size, and error range, and can automatically select a suitable statistical model according to the data characteristics.

7. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to claim 1, characterized in that: The transaction quotes and confidence intervals output by the output layer are applicable to various market models, including counterparty trading, auction trading, competitive bidding, and online matching.

8. The digital pricing technology and AI method for generating field-to-market transactions in the agricultural supply chain according to any one of claims 1-7, characterized in that, The method includes the following steps: Data collection phase: Collect market transaction characteristic data and raw metadata for fresh agricultural products, including price, output, region, user information, production information, market demand and cost data; Feature processing stage: Feature extraction, filtering, and standardization are performed through the first hidden layer of the neural system; Model optimization phase: Enhanced features are constructed through the second hidden layer, and dynamic modeling and parameter updates are performed by combining historical and similar data; Confidence interval calculation stage: Calculate the confidence interval of the price prediction results based on the statistical distribution model and Bayesian analysis; Results Output Stage: Outputs transaction quotes and their confidence ranges applicable to different trading models.

Images