CN-121998796-A - Distributed power transaction behavior identification and credit assessment system based on multi-modal neural network

CN121998796ACN 121998796 ACN121998796 ACN 121998796ACN-121998796-A

Abstract

The invention relates to the field of integration of a distributed power transaction technology and artificial intelligence, and discloses a distributed power transaction behavior recognition and credit assessment system based on a multi-mode neural network, which comprises a data acquisition layer, a data processing layer and a data processing layer, wherein the data acquisition layer is used for acquiring multi-mode time sequence data of a distributed power transaction full link in real time; the system comprises a data preprocessing layer, a multi-mode feature fusion layer, a behavior recognition and credit assessment layer, a blockchain storage and verification layer and a data trusted verification and identity authentication layer, wherein the data preprocessing layer is used for preprocessing multi-mode time sequence data to generate multi-mode sample data, the multi-mode feature fusion layer is used for extracting periodic rules, cross-mode dynamic association and long-term dependence contained in the sample data and outputting time sequence fusion features, the behavior recognition and credit assessment layer is used for carrying out transaction behavior classification recognition by adopting an intelligent optimization algorithm and calculating dynamic credit scores, and the blockchain storage and verification layer is used for storing preprocessed data, fusion features, behavior recognition results and credit scores in a blockchain and realizing data trusted verification and identity authentication. The method provides an accurate, efficient and dynamic integrated solution, and is suitable for the market trading scene of new energy.

Inventors

JIANG MING
CHEN HAO
JI CONG
CHEN MINGMING
TANG YIMING
CHEN JINING
CAI QIXIN
LIU YUNPENG
CAI MINGMING
SHAN CHAO
XIA YUHANG

Assignees

国网江苏省电力有限公司营销服务中心
江苏电力交易中心有限公司
江苏方天电力技术有限公司

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (10)

1. A distributed power transaction behavior recognition and credit assessment system based on a multi-modal neural network, comprising: The data acquisition layer is used for acquiring multi-mode time sequence data of the distributed power supply transaction all-link in real time, wherein the multi-mode time sequence data comprises transaction behavior mode data, running state mode data and trusted verification mode data; The data preprocessing layer is used for carrying out data cleaning, standardization and time sequence alignment processing based on time stamps on the multi-mode time sequence data to generate multi-mode sample data with uniform scale and synchronous time sequence; the multi-mode feature fusion layer adopts an improved SSAE-LSTM fusion architecture aiming at multi-mode sample data to realize feature extraction and cross-mode feature fusion in each mode so as to extract cycle rules, cross-mode dynamic association and long-term dependency relationship contained in the multi-mode feature fusion layer and output time sequence fusion features; The behavior recognition and credit assessment layer is used for carrying out transaction behavior classification recognition by adopting a radial basis function neural network optimized by a loop surface walking improved bat algorithm based on the time sequence fusion characteristics, outputting a behavior recognition result and calculating a credit score through a dynamic credit assessment model; The blockchain storage and verification layer is used for storing the preprocessed multi-mode time sequence data, the time sequence fusion characteristics, the behavior recognition results and the credit scores in the blockchain, and realizing the data credibility verification and the identity authentication through a consensus mechanism and an encryption algorithm.
2. The system according to claim 1, wherein the multi-modal feature fusion layer specifically comprises: the intra-mode feature extraction module adopts a first layer self-encoder of a stacked sparse self-encoder to independently train three types of mode data of transaction behavior, running state and credibility verification respectively, and extracts specific feature vectors of all modes; the cross-modal feature fusion module is used for inputting the spliced specific feature vectors of the three types of modes into a second-layer self-encoder of a stacked sparse self-encoder (SSAE) to extract cross-modal sharing features; The period decoupling and screening module is used for periodically decoupling the cross-modal sharing characteristics through fast Fourier transform to screen main period characteristics, reconstructing the main period characteristics into a three-dimensional tensor and embedding a learnable period index code; The cross-variable attention enhancement module takes the trusted verification mode feature as a query, the transaction behavior and the running state mode feature as keys and values, dynamically calculates attention weights and outputs enhanced cross-variable fusion features; and the time sequence dependency modeling module adopts a long-short-term memory network (LSTM) to perform time sequence modeling on the cross-variable fusion characteristics, captures the long-term time sequence dependency after fusion and outputs the time sequence fusion characteristics.
3. The system according to claim 2, wherein the intra-modality feature extraction module specifically performs the following process: Aiming at three types of modal data of transaction behavior, running state and credibility verification, an independent stacking sparse self-encoder is adopted for training; training of each stacked sparse self-encoder aims at minimizing a comprehensive loss function, wherein the comprehensive loss function at least comprises reconstruction errors of input data, regularization penalty terms of model weights and sparsity constraint on hidden layer neuron activation values; through training, the original time sequence data of each type of mode is encoded into a mode specific feature vector which is used as the input of a cross-mode feature fusion module.
4. The system of claim 2, wherein the period decoupling and filtering module is configured to perform frequency domain conversion and main period analysis on the cross-modal sharing feature representation, and specifically comprises: converting the cross-modal shared feature representation from the time domain to the frequency domain by fast fourier transformation, and calculating the average amplitude intensity of each frequency component in different feature dimensions; screening out a preset number of core main periods based on the average amplitude intensity; According to each screened main period, the shared characteristic representation reorganization structure is constructed into a three-dimensional tensor comprising characteristic dimension, period fragment number and period length; a learnable period index code is embedded for each period segment of the three-dimensional tensor to distinguish and identify different periodic patterns.
5. The system of claim 2, wherein the cross-variable attention enhancement module specifically performs the following: Taking the shared feature subset of the trusted verification modality as an attention query vector; the shared feature subset of the transaction behavior mode and the running state mode is used as an attention key vector and a value vector together; the method comprises the steps of dynamically distributing weights of transaction behavior and running state features in a value vector by calculating similarity of query vectors and key vectors, so that feature enhancement fusion guided by trusted verification information is realized, and enhanced cross-variable fusion features are generated and used as input of a time sequence dependent modeling module.
6. The system of claim 2, wherein the timing dependent modeling module specifically performs the following: Inputting the cross-variable fusion characteristics into a multi-layer long-short-term memory network; Screening, updating and transmitting historical information through a gating mechanism in the multi-layer long-short-term memory network so as to capture long-range dependency relationship in time sequence data; And splicing the hidden states of all the time steps to form a comprehensive time sequence fusion characteristic comprising a periodic rule, cross-variable association and long-time dependence.
7. The system of claim 1, wherein the behavior recognition and credit assessment layer comprises: The transaction behavior recognition module is used for carrying out transaction behavior classification recognition by adopting a radial basis function neural network optimized by a loop surface walking improved bat algorithm and outputting a behavior recognition result; The credit evaluation module comprises a credit evaluation model and is used for aggregating and mapping the time sequence fusion characteristics to generate a credit score of the transaction main body; and the dynamic updating module is used for jointly updating the transaction behavior identification module and the credit evaluation module by utilizing the newly generated data according to the transaction period and associating the updating result to the blockchain account.
8. The system of claim 7, wherein the transaction behavior recognition module specifically performs the following: Adopting a radial basis function neural network optimized by an improved bat algorithm as a classifier; The improved bat algorithm updates and optimizes the position of an individual by introducing chaotic inertial weight and torus walking strategy so as to balance global searching and local development capacity; the optimization objective function of the classifier synthesizes basic classification errors and cycle consistency constraints, and the cycle consistency constraints are realized by calculating characteristic differences of time sequence fusion characteristics on a main cycle; And the classifier takes the time sequence fusion characteristics output by the multi-mode characteristic fusion layer as input and outputs the recognition result of the transaction behavior class.
9. The system of claim 7, wherein the credit assessment module performs the following: Carrying out average pooling on the input time sequence fusion characteristics at all time steps to obtain a comprehensive characteristic vector; Inputting the comprehensive feature vector into a fully connected neural network for nonlinear transformation; Mapping the output of the fully connected network to an initial credit score; And carrying out weighted calculation and normalization processing on the initial credit score based on a preset score index weight system to generate a final credit score with the value range of 0 to 100, wherein the score index weight system at least comprises transaction compliance, performance rate, running stability and trusted records.
10. The system of claim 1, wherein the blockchain storage and validation layer recipe performs the following: The output data from the data preprocessing layer, the multi-mode feature fusion layer and the behavior recognition and credit assessment layer are packaged into blocks by constructing a merck tree structure, and the hash value of the block head is calculated to realize the integrity and non-tamperable storage of the data; Adopting a rights and benefits proving consensus mechanism, carrying out endorsement verification on a credit evaluation result by a preset heavy node, and confirming that the credit evaluation result is valid only when the proportion of nodes agreeing to verify reaches or exceeds a preset threshold value; The identity authentication of a transaction main body is realized through an asymmetric encryption algorithm, the transaction main body signs data by using a private key, and a network node verifies the validity of the signature by using a corresponding public key; And a data interface with an external distributed power transaction system is provided, so that credit scoring inquiry, abnormal behavior early warning and blockchain account book tracing functions are supported.

Description

Distributed power transaction behavior identification and credit assessment system based on multi-modal neural network Technical Field The invention relates to the field of integration of distributed power supply transaction technology and artificial intelligence, in particular to a distributed power supply transaction behavior identification and credit assessment system based on a multi-modal neural network. Background Along with the promotion of large-scale grid connection and power market reform of a distributed power supply, the distributed power supply transaction is faced with a plurality of technical bottlenecks due to the characteristics of more participation subjects, high transaction frequency, miscellaneous data sources and the like, firstly, transaction data cover multi-mode information such as transaction behaviors, equipment operation, credible verification and the like, each mode of data is high in heterogeneity and complex in time sequence characteristic, the traditional feature fusion method is difficult to mine cross-mode deep association, so that the behavior recognition precision is low, secondly, the dynamic association of three variables such as transaction behavior-operation state-credible verification is changed along with time in the transaction process, the static feature weighting mode cannot adapt to the dynamic coupling relation among the variables, abnormal behaviors such as malicious quotations, performance violations and the like are difficult to accurately capture, thirdly, the traditional behavior classification model is dependent on manual design features or traditional optimization algorithms, a local optimal trap exists, classification precision and generalization capability cannot be considered, meanwhile, the credit evaluation is based on static index manual scoring, dynamic linkage with real-time transaction behavior is lacked, and real-time change of the credit state of the subject is difficult to reflect. Although the application of multi-mode fusion, attention mechanisms and meta-heuristic optimization algorithms in the prior art has occurred, an integrated solution for a distributed power supply transaction scene has not been formed yet, namely the multi-mode fusion technology does not fully combine the periodic characteristics of transaction data and cross-variable correlation, the attention mechanisms do not focus on core correlation logic of 'trusted verification-transaction behavior-running state', the optimization algorithm still has defects in balance exploration and development capability, and the actual requirements of the distributed power supply transaction are difficult to meet in the aspects of behavior recognition instantaneity and credit evaluation accuracy. Therefore, it is needed to construct a transaction behavior recognition and credit assessment system integrating multi-mode deep fusion, dynamic association capturing and efficient optimization classification, so as to solve the pain point in the prior art. Disclosure of Invention The invention aims at providing an accurate, efficient and dynamic integrated solution for the technical defects in the existing distributed power supply transaction behavior recognition and credit assessment, is suitable for the market transaction scene of new energy sources such as distributed photovoltaics, wind power and the like, can realize real-time monitoring and abnormal recognition of transaction behaviors and dynamic quantitative assessment of the credit of a transaction main body, and provides technical support for compliance operation and risk management and control of an electric power transaction market. The distributed power supply transaction behavior recognition and credit assessment system based on the multi-mode neural network comprises a data acquisition layer, a data processing layer and a data processing layer, wherein the data acquisition layer is used for acquiring multi-mode time sequence data of a distributed power supply transaction all-link in real time, and the multi-mode time sequence data comprises transaction behavior mode data, running state mode data and trusted verification mode data; the system comprises a data preprocessing layer, a multi-modal feature fusion layer, a block chain storage and verification layer, a data reliability verification and identity authentication layer, a data preprocessing layer, a behavior recognition and credit assessment layer, a data reliability verification and identity authentication layer and a data reliability verification and identity authentication layer, wherein the data preprocessing layer is used for carrying out data cleaning, standardization and time-stamp-based time sequence alignment processing on multi-modal time sequence data to generate multi-modal sample data with uniform scale and synchronous time sequence, the multi-modal feature fusion layer is used for adopting an improved SSAE-LSTM fusion framework to realize feature extraction and