CN-122022892-A - Agricultural product price data opposite flushing method for electric data processing

CN122022892ACN 122022892 ACN122022892 ACN 122022892ACN-122022892-A

Abstract

The invention provides an agricultural product price data opposite flushing method for electric data processing, which is characterized in that multisource data such as agricultural product price, weather, policy, supply chain, macroscopic economy and the like are collected and fused, an evolution type dynamic causal graph is constructed by combining time sequence synchronization, standardization processing and causal relation mining, a two-channel graph neural network is introduced to realize policy generation and interpretable path synchronous reasoning, an optimal opposite flushing operation is output by using a reinforcement learning algorithm, decision logic is converted into natural language interpretation, and consistency and traceability of the policy and the interpretation are ensured.

Inventors

XIE XIAOXIN
CHEN CAI
Che Feiting
YANG DINGJUN
Huang Zide
ZHOU YUTONG

Assignees

广西兴慧教育科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. An agricultural product price data opposite flushing method for electric data processing is characterized by comprising the following steps: s1, acquiring historical price sequences of agricultural products, climate data, policy notices, supply chain logistics information and macro economic indexes as multi-source original input data; S2, performing time alignment and standardization pretreatment on the multi-source original input data, and identifying dynamic driving relations among variables to generate an initial dynamic causal graph structure; S3, taking the initial dynamic causal graph structure as a basic reasoning frame, introducing an online learning mechanism, and adjusting weight parameters and conditional probability distribution of causal edges according to market feedback data updated in real time to generate an evolution dynamic causal graph model; S4, constructing a two-channel graph neural network architecture based on the evolution type dynamic causal graph model, and introducing a synchronous attention gating mechanism to maintain the state consistency of the two channels in each reasoning step so as to form a joint decision and interpretation reasoning space; S5, in the combined decision and explanation reasoning space, performing optimal impulse searching by taking the minimum expected price risk as an objective function, generating a strategy operation sequence, and simultaneously recording a core causal path set activated in the optimal impulse searching process; and S6, carrying out semantic mapping on the core causal path set, converting the core causal path set into a logic chain description readable by natural language based on a predefined causal semantic rule base, and generating an interpretable text segment corresponding to the strategy action.
2. The method of agricultural product price data hedging by electric data processing according to claim 1, characterized in that said step S6 further comprises: s7, triggering a causal sub-graph recalibration mechanism when the logic consistency threshold requirement is not met between the strategy operation sequence and the corresponding interpretable text segment, and retrieving and optimizing the relevant causal path again based on the latest market state until the agreement is reached; And S8, outputting a final hedging scheme, wherein the final hedging scheme comprises hedging operation instructions subjected to consistency verification and complete decision basis chains matched with the hedging operation instructions, and realizing closed-loop cooperative output of strategy recommendation and interpretation generation under a unified model.
3. An agricultural product price data hedging method for electrical data processing according to claim 1, wherein nodes in said initial dynamic causal graph structure represent price fluctuation related factors, and edges represent statistically validated causal directions and intensities.
4. The method for hedging agricultural product price data by electric data processing according to claim 1, wherein said step S3 specifically comprises: based on the initial dynamic causal graph structure generated in the step S2, intercepting a recent market history data segment by utilizing a sliding time window method; Performing incremental Granges causal strength re-estimation on multi-source time sequence data in a sliding time window, and dynamically updating the directional strength coefficient of each causal edge by adopting a recursive least square method to obtain a corrected causal influence measurement matrix; Based on market feedback data flowing in real time, constructing a high-dimensional observation vector, inputting the high-dimensional observation vector to a Bayesian variation inference module, and estimating the change trend of the conditional probability distribution on each causal relationship path to obtain a change result of the conditional probability distribution; The modified causal influence measurement matrix and the conditional probability distribution change result are fused to generate a combined confidence score, and accordingly comprehensive weight parameters and node transition probabilities of all sides in the evolution type dynamic causal graph are optimized to form a graph model parameter set; and executing sparse regularization processing on the graph model parameter set, and outputting an updated dynamic causal graph.
5. The method of claim 4, wherein the market feedback data includes a range of commodity spot prices, futures contract volume transactions, policy issuing event markers, and logistic outage alert signals.
6. The method for hedging agricultural product price data by electric data processing according to claim 1, wherein said step S4 specifically comprises: constructing a bottom topological structure of a graph neural network based on a node set and a weighted directed edge set in the evolution type dynamic causal graph model to form a graph structural representation; Initializing a two-channel graph neural network architecture based on the graph structural representation, wherein the two-channel graph neural network architecture comprises the strategy generation channel and the interpretation generation channel; designing a synchronous attention gating mechanism, comparing the attention weight matrix of each reasoning step in the strategy generation channel with the corresponding attention weight matrix in the interpretation generation channel in real time, and restricting the difference between the attention weight matrix of each reasoning step in the strategy generation channel and the corresponding attention weight matrix in the interpretation generation channel to be not more than a preset threshold value through a consistency loss function; Based on the synchronized attention distribution, calculating optimal action probability distribution in the strategy generation channel, and outputting a state and action value function; And the state and action value function and the causal path liveness score are input to a decision fusion module in a combined mode, and a combined decision and interpretation reasoning space is constructed.
7. The agricultural product price data opposite flushing method based on the electric data processing of claim 6 is characterized in that the initializing two-channel graph neural network architecture is characterized in that a strategy generating channel adopts a graph attention network to conduct aggregation calculation on node characteristics, key state embedded vectors for action decision are extracted, and GAT encoders of the same structure are deployed in parallel by an interpretation generating channel, and bottom layer parameters and adjacent relations are shared.
8. The method for hedging agricultural product price data by electric data processing according to claim 1, wherein said step S5 specifically comprises: Based on node embedded vector sequences output by the evolution type dynamic causal graph model, weighting and aggregating all influence factors by using an attention mechanism, and calculating a joint characterization vector of the current market state; Inputting the joint characterization vector into a reinforcement learning framework constructed based on depth deterministic strategy gradient, and performing action space mapping to generate a continuous hedging operation suggestion sequence; In each decision time step, a synchronous activation interpretation generation channel is used for tracking a causal edge set which is obviously activated in the strategy generation process, locking a key driving path based on a synchronous attention gating mechanism, and generating a corresponding intermediate causal track sequence; performing path pruning and importance sequencing on the intermediate causality track sequence, and screening a core causality path set with the contribution degree exceeding a preset standard to the current strategy action by utilizing a causality intensity threshold; and collecting the continuous hedging operation suggestion sequence and the corresponding core causal path as a joint output result.
9. The method of claim 8, wherein the continuous hedging operation recommendation sequence includes futures binning direction, option combination configuration ratio, and spot modulation range.
10. An agricultural product price data hedging method for electric data processing according to claim 1, characterized in that said interpretable text segment contains a complete logic chain for impulse identification, trigger conditions, core drivers, intermediate conduction paths and final influence inference.

Description

Agricultural product price data opposite flushing method for electric data processing Technical Field The invention relates to the technical field of market data analysis and intelligent decision support, in particular to an agricultural product price data opposite flushing method for electric data processing. Background The existing agricultural product price risk management and hedging strategy recommendation technology mainly adopts price data driving, and models such as machine learning, time sequence analysis, expert rules and the like are adopted to model and predict market fluctuation. The main stream method is to extract the characteristics of historical price, climate, policy information and supply chain index, and on the basis, adopt a deep neural network, regression model or integrated learning method to generate a hedging scheme, and assist a knowledge graph, association rules or causal network to promote the interpretability. Part of the technical system tries to realize knowledge-driven intelligent decision-making; In the prior art, a strategy recommendation module and a strategy interpretation module are usually realized separately, a recommendation flow is focused on outputting an optimal opposite impulse action sequence, and the interpretation module carries out traceability description on a scheme result independently according to a knowledge graph or a causality relation. The structure causes a logic fault between the interpretation content and the strategy recommendation, and cannot ensure the strict causal alignment of the interpretation text and the actual recommendation action, thereby affecting the integral credibility and traceability of the system; Typical representative technologies such as agricultural product price prediction models based on time series analysis, market trend capture through ARIMA, LSTM or VAR models are suitable for price trend prediction and conventional risk management and control scenes. The prior art also provides an interpretation enhancement technology based on causal inference and a knowledge graph, which can assist users to understand recommendation results and decision bases thereof to a certain extent, but often carries out post-hoc interpretation on the existing decisions after recommendation, and cannot realize real-time synchronization of strategy actions and interpretation paths. Disclosure of Invention The invention aims to solve the technical problems and provides an agricultural product price data opposite flushing method for electric data processing. The technical scheme of the invention is realized in such a way that the agricultural product price data opposite flushing method for electric data processing comprises the following steps: S1, acquiring historical price sequences of agricultural products, climate data, policy announcements, supply chain logistics information and macro economic indexes as multi-source original input data so as to construct an external influence factor set generated by an opposite strategy; S2, performing time alignment and standardization pretreatment on the multi-source original input data, and identifying dynamic driving relations among variables based on a Granges causal test and a time sequence nonlinear dependency analysis method to generate an initial dynamic causal graph structure, wherein nodes represent price fluctuation related factors, and edges represent causal directions and intensities verified by statistics; S3, taking the initial dynamic causal graph structure as a basic reasoning frame, introducing an online learning mechanism, and adjusting weight parameters and conditional probability distribution of causal edges according to market feedback data updated in real time to generate an evolution dynamic causal graph model with environment adaptability, wherein the evolution dynamic causal graph model is used for representing a price influence mechanism which changes along with time; S4, constructing a dual-channel graph neural network architecture based on the evolution type dynamic causal graph model, wherein a strategy generation channel and an interpretation generation channel share the same graph structural representation, and ensuring that the two channels maintain state consistency in each reasoning step through a synchronous attention gating mechanism to form a joint decision-interpretation reasoning space; s5, in the joint decision-interpretation reasoning space, performing optimal impulse searching by using a reinforcement learning algorithm with the minimum expected price risk as an objective function to generate a strategy operation sequence, and simultaneously recording a core causal path set activated in the process; S6, inputting the core causal path set into a semantic mapping module, converting the core causal path set into a logic chain description readable by natural language based on a predefined causal semantic rule base, and generating an interpretable text segment strictly corr