Distributed power supply transaction behavior recognition and credit evaluation system based on multi-modal neural network

Abstract

The invention relates to the field of distributed power supply transaction technology and artificial intelligence fusion, and discloses a distributed power supply transaction behavior recognition and credit evaluation system based on a multi-modal neural network, and the system comprises a data collection layer which collects the multi-modal time sequence data of a distributed power supply transaction full link in real time; the data preprocessing layer is used for preprocessing the multi-modal time sequence data to generate multi-modal sample data; the multi-modal feature fusion layer is used for extracting a periodic rule, cross-modal dynamic association and a long-time-history dependency relationship contained in the sample data and outputting time sequence fusion features; the behavior recognition and credit evaluation layer is used for carrying out transaction behavior classification recognition by adopting an intelligent optimization algorithm and calculating a dynamic credit score; and the block chain storage and verification layer is used for storing the preprocessed data, the fusion features, the behavior recognition result and the credit score in a block chain and realizing data credibility verification and identity authentication. And an accurate, efficient and dynamic integrated solution is provided, and the method is suitable for a market transaction scene of new energy.

Description

Technical Field

[0001] This invention relates to the field of distributed power trading technology and artificial intelligence integration, and in particular to a distributed power trading behavior recognition and credit assessment system based on multimodal neural networks. Background Technology

[0002] With the large-scale grid connection of distributed power sources and the advancement of power market reform, distributed power source trading faces numerous technical bottlenecks due to the large number of participants, high trading frequency, and diverse data sources. Firstly, trading data encompasses multimodal information such as trading behavior, equipment operation, and trusted verification. The data exhibits strong heterogeneity and complex temporal characteristics, making it difficult for traditional feature fusion methods to uncover deep cross-modal correlations, resulting in low behavior recognition accuracy. Secondly, the dynamic correlation between the three variables—"trading behavior, operating status, and trusted verification"—changes over time during the trading process. Static feature weighting methods cannot adapt to the dynamic coupling relationships between variables, making it difficult to accurately capture abnormal behaviors such as malicious bidding and contract default. Thirdly, existing behavior classification models often rely on manually designed features or traditional optimization algorithms, which are prone to local optima traps and cannot balance classification accuracy and generalization ability. Furthermore, credit assessments are often based on static indicators and manual scoring, lacking dynamic linkage with real-time trading behavior and failing to reflect real-time changes in the credit status of entities.

[0003] Although existing technologies have incorporated multimodal fusion, attention mechanisms, and metaheuristic optimization algorithms, a unified solution for distributed power generation trading scenarios has yet to be developed. Multimodal fusion technology has not fully integrated the periodic characteristics and cross-variable correlations of trading data; attention mechanisms have not focused on the core correlation logic of "trustworthy verification - trading behavior - operating status"; and optimization algorithms still fall short in balancing exploration and development capabilities. As a result, existing systems are unable to meet the actual needs of distributed power generation trading in terms of real-time behavior recognition and accuracy of credit assessment.

[0004] Therefore, there is an urgent need to build a transaction behavior recognition and credit assessment system that integrates multimodal deep fusion, dynamic correlation capture, and efficient optimized classification to solve the pain points of existing technologies. Summary of the Invention

[0005] The core objective of this invention is to address the technical deficiencies in existing distributed power generation trading behavior identification and credit assessment by providing a precise, efficient, and dynamic integrated solution applicable to market-based trading scenarios for distributed photovoltaic, wind power, and other new energy sources. This solution enables real-time monitoring of trading behavior, anomaly identification, and dynamic quantitative assessment of the creditworthiness of trading entities, providing technical support for compliant operation and risk management in the power trading market.

[0006] To achieve the above objectives, the following technical solution is provided: This invention provides a distributed power supply transaction behavior recognition and credit assessment system based on a multimodal neural network, comprising: a data acquisition layer for real-time acquisition of multimodal time-series data across the entire distributed power supply transaction chain, including: transaction behavior modal data, operational status modal data, and trusted verification modal data; a data preprocessing layer for data cleaning, standardization, and timestamp-based time-series alignment of the multimodal time-series data to generate multimodal sample data with uniform scale and synchronized time series; and a multimodal feature fusion layer for employing an improved SSAE-LSTM fusion framework for the multimodal sample data. The system comprises a structure that enables feature extraction within each modality and cross-modal feature fusion to extract the inherent periodic patterns, cross-modal dynamic correlations, and long-term dependencies, outputting time-series fusion features; a behavior recognition and credit assessment layer that, based on the time-series fusion features, uses a radial basis function neural network optimized by a toroidal walking improved bat algorithm to classify and recognize transaction behaviors, outputting behavior recognition results, and calculating credit scores through a dynamic credit assessment model; and a blockchain storage and verification layer that stores the preprocessed multimodal time-series data, time-series fusion features, behavior recognition results, and credit scores on the blockchain, and achieves data credibility verification and identity authentication through consensus mechanisms and encryption algorithms.

[0007] Furthermore, the multimodal feature fusion layer specifically includes: an intramodal feature extraction module, which uses the first layer of a stacked sparse autoencoder (SSAE) to independently train the three modalities of transaction behavior, running state, and trusted verification data, and extracts the specific feature vectors of each modality; a cross-modal feature fusion module, which concatenates the specific feature vectors of the three modalities and inputs them into the second layer of the stacked sparse autoencoder (SSAE) to extract cross-modal shared features; a periodic decoupling and filtering module, which uses a fast Fourier transform to periodically decouple the cross-modal shared features to filter the main periodic features, and reconstructs them into a three-dimensional tensor, embedding a learnable periodic index code; a cross-variable attention enhancement module, which uses trusted verification modal features as queries and transaction behavior and running state modal features as keys and values, dynamically calculates attention weights, and outputs the enhanced cross-variable fusion features; and a temporal dependency modeling module, which uses a long short-term memory network (LSTM) to perform temporal modeling on the cross-variable fusion features, captures the long-term temporal dependencies of the fused features, and outputs the temporal fusion features.

[0008] Furthermore, the intra-modal feature extraction module performs the following process: for the three modalities of transaction behavior, running status, and trust verification, independent stacked sparse autoencoders are trained respectively; the training of each stacked sparse autoencoder aims to minimize the comprehensive loss function, which includes at least the reconstruction error of the input data, the regularization penalty term of the model weights, and the sparsity constraint on the activation values ​​of the hidden layer neurons; through training, the original time-series data of each modality is encoded into a modality-specific feature vector, which is used as the input of the cross-modal feature fusion module.

[0009] Furthermore, the periodic decoupling and filtering module is configured to perform frequency domain transformation and principal period analysis on the cross-modal shared feature representation, specifically including: transforming the cross-modal shared feature representation from the time domain to the frequency domain through fast Fourier transform, and calculating the average amplitude intensity of each frequency component in different feature dimensions; filtering out a preset number of core principal periods based on the average amplitude intensity; reorganizing the shared feature representation according to each selected principal period to construct a three-dimensional tensor containing feature dimensions, number of periodic segments, and period length; embedding a learnable periodic index code for each periodic segment of the three-dimensional tensor to distinguish and identify different periodic patterns.

[0010] Furthermore, the cross-variable attention enhancement module performs the following process: using a subset of shared features from the trusted verification modality as the attention query vector; using a subset of shared features from the transaction behavior modality and the running state modality as the attention key vector and value vector; and dynamically allocating the weights of transaction behavior and running state features in the value vector by calculating the similarity between the query vector and the key vector, thereby achieving feature enhancement fusion guided by trusted verification information and generating enhanced cross-variable fusion features as input to the temporal dependency modeling module.

[0011] Furthermore, the time-series dependency modeling module is configured to: input the cross-variable fusion features into a multi-layer long short-term memory network; through the gating mechanism in the multi-layer long short-term memory network, realize the filtering, updating and transmission of historical information to capture long-range dependencies in time-series data; and splice the hidden states of all time steps to form a comprehensive time-series fusion feature that includes periodic patterns, cross-variable associations and long-term dependencies.

[0012] Furthermore, the behavior recognition and credit assessment layer includes: a transaction behavior recognition module, which uses a radial basis function neural network optimized by a toroidal walking improved bat algorithm to classify and recognize transaction behaviors and outputs behavior recognition results; a credit assessment module, which includes a credit assessment model for aggregating and mapping the temporal fusion features to generate a credit score for the transaction entity; and a dynamic update module, which uses newly generated data to jointly update the transaction behavior recognition module and the credit assessment module according to the transaction cycle and associates the update results with the blockchain account.

[0013] Furthermore, the transaction behavior recognition module specifically executes the following process: a radial basis function neural network optimized by an improved bat algorithm is used as the classifier; the improved bat algorithm updates and optimizes the position of individuals by introducing chaotic inertial weights and a toroidal walking strategy to balance global search and local exploitation capabilities; the optimization objective function of the classifier integrates the basic classification error and the periodic consistency constraint, which is achieved by calculating the feature differences of the temporal fusion features in the main period; the classifier takes the temporal fusion features output by the multimodal feature fusion layer as input and outputs the recognition result of the transaction behavior category.

[0014] Furthermore, the credit assessment module is configured to perform the following process: average pooling is performed on the input temporal fusion features at all time steps to obtain a comprehensive feature vector; the comprehensive feature vector is input into a fully connected neural network for nonlinear transformation; the output of the fully connected network is mapped to an initial credit score; based on a preset scoring indicator weighting system, the initial credit score is weighted and normalized to generate a final credit score with a value range of 0 to 100; wherein, the scoring indicator weighting system includes at least transaction compliance, fulfillment rate, operational stability, and reliable records.

[0015] Furthermore, the blockchain storage and verification layer performs the following process: Output data from the data preprocessing layer, multimodal feature fusion layer, and behavior recognition and credit assessment layer are encapsulated into blocks by constructing a Merkle tree structure, and the block header hash value is calculated to achieve data integrity and tamper-proof storage; a proof-of-stake consensus mechanism is adopted, with preset key nodes endorsing and verifying the credit assessment results. Only when the proportion of nodes agreeing to verification reaches or exceeds a preset threshold is the credit assessment result confirmed as valid; identity authentication of the transaction entity is achieved through an asymmetric encryption algorithm, with the transaction entity using its private key to sign the data, and network nodes using the corresponding public key to verify the signature validity; a data interface with an external distributed power trading system is provided, supporting credit score queries, abnormal behavior warnings, and blockchain ledger traceability functions.

[0016] Compared with the prior art, the present invention achieves the following beneficial effects:

[0017] 1. This invention constructs an improved SSAE-LSTM fusion architecture, which deeply mines the intrinsic correlation of three types of multimodal data: transaction behavior, running status, and trusted verification by extracting sparse features within a modality and fusing common features across modalities. At the same time, it combines the LSTM gating mechanism to capture long-term temporal dependencies, thus solving the problem of insufficient feature representation caused by the heterogeneity and temporal complexity of multimodal data.

[0018] 2. This invention addresses the dynamic correlation characteristics of three types of variables in distributed power generation transactions: "transaction behavior - operating status - trusted verification". It introduces a cross-variable attention enhancement module (CvAM), which uses trusted verification modal features as queries (Q), transaction behavior and operating status modal features as keys (K) and values ​​(V), and dynamically allocates variable weights to accurately capture the coupling relationship of the three types of variables over time, thereby improving the sensitivity of abnormal behavior identification.

[0019] 3. This invention constructs an RBF neural network classifier optimized by the toroidal walking improved bat algorithm (TW-BA). By exploring and developing the chaotic inertia weight balancing algorithm and combining it with the toroidal walking strategy to expand the search space, it solves the problem that traditional optimization algorithms are prone to getting trapped in local optima, and improves the accuracy and generalization ability of transaction behavior classification.

[0020] 4. This invention establishes a dynamic credit assessment system linked to real-time behavior recognition, realizes the periodic updating of credit scores, and relies on blockchain technology to ensure data credibility and the immutability of assessment results, providing fair and transparent credit support for the distributed power trading market.

[0021] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0022] in:

[0023] Figure 1 This is a schematic diagram of the modules of the distributed power trading behavior recognition and credit assessment system based on multimodal neural networks provided in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the system architecture of the distributed power trading behavior recognition and credit assessment system based on multimodal neural networks provided in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the improved SSAE-LSTM fusion architecture according to an embodiment of the present invention;

[0026] Figure 4This is a schematic diagram of the structure of the RBF neural network (TW-BA-RBF) optimized by the improved bat algorithm for toroidal walking according to an embodiment of the present invention;

[0027] Figure 5 This is a power trading flowchart according to an embodiment of the present invention;

[0028] Figure 6 This is a screenshot of data sent by a blockchain node in an embodiment of the present invention;

[0029] Figure 7 This is the statistical result of blockchain feedback data in the embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0031] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a distributed power trading behavior recognition and credit assessment system 100 based on a multimodal neural network, including a data acquisition layer 110, a data preprocessing layer 120, a multimodal feature fusion layer 130, a behavior recognition and credit assessment layer 140, and a blockchain storage and verification layer 150 connected in sequence and coordinating data.

[0032] The data acquisition layer 110 is used to collect multimodal time-series data of the entire distributed power trading chain in real time, including: trading behavior modal data, operation status modal data and trusted verification modal data.

[0033] Step 1: Construct the data acquisition layer 110:

[0034] Collect multi-modal time-series data of the entire distributed power generation transaction chain, including three core modes:

[0035] (1) Transaction behavior mode: Real-time collection of transaction declaration data (declared electricity volume, electricity price, transaction time period), transaction data (transaction electricity volume, execution price, performance progress), and account operation data (login time, declaration frequency, modification record) through blockchain nodes to form a time-series transaction sequence;

[0036] (2) Operational status mode: Collect distributed power generation unit operation data (output power, equipment load rate, fault records) and grid dispatch response data (access stability, load matching degree), and generate time-series operation sequences synchronously according to the transaction cycle;

[0037] (3) Trusted verification mode: Extract immutable data such as the identity authentication information of the transaction subject, the hash value of historical performance records, the status of margin payment, and the node endorsement records from the blockchain ledger to form a trusted time-series verification sequence.

[0038] The data preprocessing layer 120 is used to perform data cleaning, standardization and time-stamp-based time-series alignment on multimodal time-series data to generate multimodal sample data with uniform scale and synchronized time series.

[0039] Step 2: Construct the data preprocessing layer 120

[0040] Step 2.1: Data Cleaning

[0041] Box plot anomaly detection combined with linear interpolation is used to handle missing values ​​and outlier data, and first-order exponential smoothing is used to eliminate high-frequency noise in the running status data.

[0042] Step 2.2: Data Standardization

[0043] z-score normalization (transaction behavior mode) and min-max normalization (running state mode) are performed on the three types of modal data respectively to ensure data scale consistency.

[0044] Step 2.3: Timing Alignment

[0045] Based on blockchain timestamps, multimodal data with different sampling frequencies are uniformly aligned to 5 minutes / step to generate time series samples of equal duration. Sample labels include normal / abnormal transaction behavior labels and historical credit rating labels.

[0046] Multimodal feature fusion layer 130 adopts an improved SSAE-LSTM fusion architecture for multimodal sample data to realize feature extraction within each modality and cross-modal feature fusion, so as to extract the periodic patterns, cross-modal dynamic correlations and long-term dependencies contained therein, and output temporal fusion features;

[0047] Furthermore, the multimodal feature fusion layer 130 specifically includes:

[0048] The intramodal feature extraction module 131 uses the first layer of the stacked sparse autoencoder (SSAE) to independently train the three modal data of transaction behavior, running status and trust verification, and extract the specific feature vectors of each modality.

[0049] The cross-modal feature fusion module 132 concatenates the specific feature vectors of the three modalities and inputs them into the second layer of the stacked sparse autoencoder (SSAE) to extract cross-modal shared features.

[0050] The periodic decoupling and filtering module 133 performs periodic decoupling on cross-modal shared features through fast Fourier transform to filter the main periodic features and reconstructs them into a three-dimensional tensor, embedding a learnable periodic index code.

[0051] The cross-variable attention enhancement module 134 uses trusted verification modal features as queries, transaction behavior and running status modal features as keys and values, dynamically calculates attention weights, and outputs enhanced cross-variable fusion features.

[0052] The temporal dependency modeling module 135 uses a long short-term memory network (LSTM) to perform temporal modeling on cross-variable fusion features, capture the long-term temporal dependencies after fusion, and output the temporal fusion features.

[0053] The multimodal feature fusion layer 130 is constructed more specifically through the following three steps.

[0054] Step 3: Construct a multimodal feature fusion layer

[0055] Figure 3 This is a schematic diagram of the improved SSAE-LSTM fusion architecture according to an embodiment of the present invention, as shown below. Figure 3 As shown, an improved SSAE-LSTM fusion architecture is adopted to achieve deep mining of cross-modal features and capture of temporal dependencies.

[0056] Step 3.1 Intramodal Feature Extraction

[0057] The intra-modal feature extraction module 131 performs the following process: for the three types of modal data—transaction behavior, running status, and trust verification—independent stacked sparse autoencoders are trained respectively; the training of each stacked sparse autoencoder aims to minimize the comprehensive loss function, which includes at least the reconstruction error of the input data, the regularization penalty term of the model weights, and the sparsity constraint on the activation values ​​of the hidden layer neurons; through training, the original time-series data of each modality is encoded into a modality-specific feature vector, which is used as the input of the cross-modal feature fusion module 132.

[0058] The specific implementation is as follows: The first layer of a stacked sparse autoencoder (SSAE) is used to independently train the three modalities. KL divergence sparse penalty and weight decay regularization are introduced to mine high-order correlation features within each modality. Taking the trading behavior modality as an example, its loss function is:

[0059]

[0060]

[0061] The intra-modal feature extraction loss function for the trading behavior modality is used to optimize the training of the modality's autoencoder and measures the sum of reconstruction error, weight regularization loss, and sparsity penalty loss. The time step of time series data refers to the number of time series observation points contained in a single sample (e.g., the number of daily sampling points under a 5-minute / step). The reconstructed output of the transaction behavior modality is generated by the autoencoder decoding layer and is used to compare with the original aligned data to calculate the reconstruction error. : Time-aligned transaction behavior modal data, and preprocessed (cleaning, standardization, alignment) input data. : The square of the Frobenius norm, used to calculate the reconstruction error of matrix-form data (time series feature matrix). The weight decay coefficient for the trading behavior modality is used to suppress overfitting of the autoencoder weight matrix, and its value range is typically [value range missing]. . : The weight matrix of the transaction behavior modal autoencoder, and the parameter matrix connecting the input layer and the hidden layer. : The sparsity penalty coefficient for the transaction behavior modality, used to control the sparsity of neurons in the hidden layer, and its value typically ranges from 0 to 1. . The number of neurons in the first hidden layer of SSAE, used to extract higher-order features within a modality. KL divergence is used to measure the difference between the average activation value of neurons in the hidden layer and the preset sparsity, thereby achieving sparsity regularization. The preset sparsity of the trading behavior modality, i.e., the expected average activation probability of neurons in the hidden layer, is typically set to a value of [value missing]. . : Transaction Behavior Modal Autoencoder The average activation value of a hidden layer neuron is obtained by averaging the activation values ​​of that neuron over all time steps. : Transaction Behavior Modal Autoencoder The hidden layer neurons in the first... The activation values ​​at each time step are calculated using an activation function (such as sigmoid or ReLU).

[0062] Similarly, the running state mode and the trusted verification mode are trained through their own intramodal autoencoders, outputting mode-specific feature vectors. .

[0063] : The specific feature vector of the transaction behavior mode, output by the first autoencoder layer, contains the high-order unique features of this mode. : Specific feature vectors of the operating state mode, and The structure is consistent, corresponding to the high-order unique features of the operating state mode. : Specific feature vectors of trusted verification modes, and The structure is consistent, corresponding to the high-order unique features of the trusted verification modality. The data dimension representation of the feature vector. For time step, The number of neurons in the first hidden layer is denoted as , which is the feature vector. OK A matrix of columns.

[0064] Step 3.2 Cross-modal feature fusion

[0065] The specific feature vectors of the three modalities are concatenated into a fusion input. Cross-modal associations are trained using the second layer autoencoder of SSAE, with the loss function being:

[0066]

[0067] Cross-modal feature fusion loss function, used to optimize the training of the second-layer autoencoder of SSAE, measures the sum of cross-modal reconstruction error, weight regularization loss and sparsity penalty loss. The reconstructed output after cross-modal fusion is generated by the second autoencoder decoding layer and is used to compare with the spliced ​​fusion input to calculate the reconstruction error. The input is a concatenation of three modality-specific feature vectors, i.e. It includes all unique features of transaction behavior, operational status, and trusted verification modality. : Fusion Input Data dimension representation, It is the sum of the three modality-specific feature dimensions (each modality feature dimension is 1). ). : Weight decay coefficient for cross-modal fusion, used to suppress overfitting of the weight matrix of the second-layer autoencoder, and its value range is usually [0.0001, 0.001]. : The weight matrix of the cross-modal autoencoder, and the parameter matrix connecting the fused input layer and the second hidden layer. : The sparsity penalty coefficient for cross-modal fusion, used to control the sparsity of neurons in the second hidden layer, typically ranging from [value range missing]. . The number of neurons in the second hidden layer of SSAE, used to extract cross-modal shared features. The preset sparsity for cross-modal fusion is the expected average activation probability of neurons in the second hidden layer, typically set to a value of [value missing]. . Cross-modal autoencoder The average activation value of a hidden layer neuron is obtained by averaging the activation values ​​of that neuron over all time steps.

[0068] Output a shared feature representation with uniform dimension. . Cross-modal shared feature representation, output by the second autoencoder layer, contains common high-order features of three modalities, realizing a unified feature representation of multimodal data. Data dimension representation with shared feature representation, For time step, This represents the number of neurons in the second hidden layer.

[0069] Step 3.3 Periodic Decoupling and Filtering

[0070] Periodic decoupling and filtering module 133: Performs frequency domain transformation and principal period analysis on the cross-modal shared feature representation, specifically including: transforming the cross-modal shared feature representation from the time domain to the frequency domain through Fast Fourier Transform, and calculating the average amplitude intensity of each frequency component in different feature dimensions; filtering out a preset number of core principal periods based on the average amplitude intensity; reorganizing the shared feature representation according to each selected principal period to construct a three-dimensional tensor containing feature dimensions, number of periodic segments, and period length; embedding a learnable periodic index code into each periodic segment of the three-dimensional tensor to distinguish and identify different periodic patterns. The specific implementation is as follows:

[0071] Shared features are represented using Fast Fourier Transform (FFT). Convert to the frequency domain and calculate the amplitude of each frequency component:

[0072]

[0073] FFT: Fast Fourier Transform, used to represent shared features in the time domain. By converting to the frequency domain, we can uncover periodic patterns in time-series data. Shared features represent the frequency domain signal after FFT transformation, which contains multiple frequency components. Frequency domain signal The real part of the frequency domain features constitutes the real part of the frequency domain features. Frequency domain signal The imaginary part constitutes the imaginary part of the frequency domain characteristics. : No. The amplitude of each frequency component is calculated from the real and imaginary parts of the frequency domain signal, reflecting the intensity of that frequency component. : The th under all feature dimensions The average amplitude of each frequency component is used to comprehensively assess the importance of that frequency component. : Averaging function, used to calculate the mean of the amplitudes of multiple frequency components, here it is... The mean amplitude corresponding to each feature dimension. : No. The th feature dimension The amplitude of each frequency component For shared features The feature dimensions.

[0074] Based on the average amplitude intensity Sort all frequency components in descending order of their average amplitude, and select the top ones. The period corresponding to the frequency component with the largest amplitude is taken as the core principal period. ,Will Reconstructed A two-dimensional tensor ( For the number of segments, (where the period length is 1), and the learned periodic index code (CIE) is embedded to capture pseudo-periodic features. : Filtered One main cycle, The preset number of master cycles (usually set to a value) This reflects the core time series period (such as daily period, weekly period) of distributed power generation transaction data. : No. The three-dimensional tensor corresponding to each principal period is composed of shared features. The reconstructed data is used to encapsulate the long-term and short-term variation characteristics within this cycle. : 3D tensor Data dimension representation, To share feature dimensions, This represents the number of segments (number of long-term changes) within that period. The period length (number of short-term change steps). : No. The number of segments in a main cycle, i.e. the number of times that cycle repeats in time series data, reflects the long-term variation pattern. : No. The length of each main period, i.e., the time step contained in a single period, reflects short-term variation patterns. CIE: Learnable Cycle Index Encoding, assigns a unique identifier to each pseudo-period segment to help the model distinguish features of different periods.

[0075] Step 3.4: Enhanced cross-variable attention

[0076] To address the technical bottleneck in the background technology where "static feature weighting methods cannot adapt to the dynamic coupling relationships between variables," this invention introduces a cross-variable attention enhancement module (CvAM). Its core design logic lies in the fact that in distributed power trading, the credibility of anomalies in transaction behavior and equipment operating status ultimately needs to be determined by trusted verification information (such as identity and performance records) on the blockchain. The cross-variable attention enhancement module 134 uses "trusted verification modal features as the query (Q)" to actively retrieve and measure the importance of transaction behavior and operating status modal features (K and V). Through this "trustworthiness"-oriented dynamic attention allocation mechanism, the model can adaptively focus on transaction and operating clues most relevant to the current trustworthiness state, thereby accurately capturing the coupling relationships that change over time and achieving highly sensitive identification of abnormal behaviors such as malicious bidding and performance default.

[0077] The cross-variable attention enhancement module 134 specifically executes the following process: It uses a subset of shared features from the trusted verification modality as the attention query vector; it uses a subset of shared features from the transaction behavior modality and the running state modality together as the attention key vector and value vector; by calculating the similarity between the query vector and the key vector, it dynamically allocates the weights of transaction behavior and running state features in the value vector, thereby achieving feature enhancement fusion guided by trusted verification information, generating enhanced cross-variable fusion features, which serve as input to the temporal dependency modeling module 135. The specific implementation is as follows:

[0078] To address the dynamic correlation among three types of variables—"transaction behavior," "operational status," and "trusted verification"—in distributed power generation transactions, a cross-variable attention enhancement module (CvAM) is introduced, using trusted verification modal features as the query ( The modal characteristics of transaction behavior and operational state are defined as keys (K) and values ​​(V), with variable weights dynamically assigned:

[0079]

[0080]

[0081] Attention Query, derived from the shared features of the trusted verification modality, is used to query the association between the modality and other modalities. Attention key (Key) is obtained by concatenating the shared features of transaction behavior and runtime mode, and is used to calculate attention weights for matching with queries. Attention value (Value), and Homologous features are obtained by concatenating shared features of transaction behavior and operational state modes, and are used for feature weighted fusion based on attention weights. : Query the linear transformation matrix, used to map the shared features of trusted verification modalities to the query space, with dimensions of ( (For key dimensions). The key linear transformation matrix is ​​used to map the concatenated bimodal shared features to the key space, with dimensions of . . : Value linear transformation matrix, used to map the concatenated bimodal shared features to the value space, with dimension . . A subset of shared features of trusted verification modalities, from The features obtained from the separation process are focused on trustworthy and relevant characteristics. A subset of shared features of transaction behavior modalities, from The data is separated from the data and focuses on transaction-related features. : A shared feature subset of the running state modes, from The features obtained from the separation process are focused on operational characteristics. : The concatenation operation of shared features between transaction behavior and operational state modalities, with dimensions of Attention : A cross-variable attention computation function used to dynamically capture the correlation strength between the trusted verification modality and the other two modalities, outputting enhanced fusion features. Softmax The Softmax activation function is used to normalize the attention score into weights (within a range of values). ), ensuring that the sum of the weights is 1. : Matrix multiplication of queries and keys, calculating the match score of each query with all keys, with dimension . . The square root of the key dimension is used to scale the attention score, avoiding numerical overflow or gradient vanishing due to excessive dimensionality. Key dimension, i.e. and The feature dimension typically takes values ​​of and . Consistency ensures that matrix multiplication can be performed.

[0082] , , To share feature representations across modalities The feature subset separated from the data. In a preferred embodiment, the feature subset is... In the feature dimension (i.e.) It is divided into three parts in terms of dimension, such that: , , .in, This indicates rounding down, and this equal division reflects equal treatment of the three modalities and provides a foundation for subsequent attention calculations. Those skilled in the art will understand that other methods (such as using different linear transformation matrices) can also be used to achieve this. Extract the feature subset.

[0083] Output Enhanced Cross-Variable Fusion Features . The enhanced cross-variable fusion feature integrates the guidance information of the trusted verification modality with key features from the other two modalities, with a dimension of [missing information]. . : Data dimension representation of cross-variable fusion features For time step, For feature dimensions.

[0084] Step 3.5: Temporal Dependency Modeling

[0085] Temporal Dependency Modeling Module 135: Fusing Features Across Variables The data is input into a multi-layered long short-term memory (LSTM) network. Through the gating mechanism within the LSM network, historical information is filtered, updated, and transmitted to capture long-range dependencies in the time-series data. The hidden states of all time steps are concatenated to form a comprehensive time-series fusion feature that includes periodic patterns, cross-variable correlations, and long-term dependencies. The specific implementation is as follows:

[0086] Cross-variable fusion features Input a two-layer LSTM network and use a gating mechanism to capture long-term time series dependencies. The calculation of each LSTM gate is as follows:

[0087] Forgotten Gate: ;

[0088] Input Gate: ;

[0089] Cell status update: ;

[0090] Output gate: ;

[0091] Hidden layer output: .

[0092] Cross-variable fusion features, used as input to the LSTM network, include periodic features and cross-variable association features. Forget gate, output range of values This is used to determine whether to retain the cell state information from the previous moment (1 means to retain it completely, and 0 means to forget it completely). : Input gate, output value range [0,1], used to determine whether the input features at the current time are updated to the cell state. : Candidate cell state, calculated from the current input and the hidden layer output of the previous time step, containing new feature information of the current time step. Cellular state, the core memory unit of LSTM, is updated through forget gate and input gate to preserve long-term temporal dependency information. Output gate, output value range It is used to determine which information in the cell state is output to the hidden layer. LSTM in the The hidden layer output at each time step contains a temporal feature representation of the current moment. LSTM in the The hidden layer output at each time step conveys the feature information from the previous time step. Shared feature representation in the first The feature vectors at each time step are replaced here with... (Intervariate fusion features in the first) (a vector at each time step). The weight matrix of the LSTM corresponds to the linear transformation parameters of the forget gate, input gate, candidate cell state, and output gate, respectively. The bias vectors of the LSTM correspond to the bias parameters of the forget gate, input gate, candidate cell state, and output gate, respectively. : sigmoid activation function, range of values It is used to control the degree of opening and closing of the gate. Hyperbolic tangent activation function, range of values , used to perform nonlinear transformations on candidate cell states and cell states, enhancing feature representation capabilities. ⊙: Element-wise product operation (Hadamard product), used to perform element-wise multiplication of two matrices / vectors of the same dimension, achieving gating for information filtering.

[0093] Final output temporal fusion features . The temporal fusion feature is obtained by concatenating the outputs of the LSTM hidden layers at all time steps, and includes comprehensive features such as periodicity, cross-variable correlation, and long-term temporal dependence. : Data dimension representation of time-series fusion features For time step, The number of neurons in the LSTM hidden layer (usually set to a value of...) This determines the ability of LSTM to capture timing dependencies.

[0094] In summary, the improved SSAE-LSTM fusion architecture of this invention is not a simple model stacking, but rather achieves deep fusion through a cascaded structure of "feature compression-temporal unrolling". Specifically:

[0095] (1) Structural connection relationship: The SSAE module, as the front-end feature extractor, ultimately outputs the shared feature representation. Instead of directly inputting into the LSTM, the data first passes through a cross-variable attention enhancement module (CvAM) to obtain enhanced cross-variable fusion features. , The feature vector integrates periodic information and cross-modal dynamic correlation, serving as the input to the LSTM network at each time step.

[0096] (2) Data Dimension Transformation: Features after SSAE and CvAM processing It is a dimension The matrix, where T is the time step, This is the feature dimension. The matrix is ​​sliced ​​by time step such that at each time step from t=1 to t=T, the input to the LSTM network is... That is, one The LSTM network is responsible for learning the dynamic evolution of these eigenvectors over time.

[0097] (3) Difference from ordinary concatenation: The architecture of this invention is different from simply connecting the SSAE output to LSTM. The key is that a CvAM module and a periodic decoupling link are introduced in the middle, so that the features input to LSTM have been dynamically associated and periodically structured, so that LSTM can focus more on learning pure, high-quality temporal dependencies.

[0098] Behavior recognition and credit assessment layer 140, based on temporal fusion features A radial basis function neural network optimized by a toroidal walking improved bat algorithm is used to classify and identify transaction behavior, output the behavior identification results, and calculate the credit score through a dynamic credit assessment model.

[0099] Step 4: Constructing the Behavior Recognition and Credit Assessment Layer 140

[0100] Step 4.1: Transaction Behavior Identification

[0101] The behavior recognition and credit assessment layer 140 includes a transaction behavior recognition module 141, which uses a radial basis function neural network optimized by an improved bat algorithm with toroidal walking to classify and recognize transaction behaviors, and outputs the behavior recognition results. The transaction behavior recognition module 141 is configured to specifically execute the following process: using a radial basis function neural network optimized by an improved bat algorithm as a classifier; the improved bat algorithm updates and optimizes the individual's position by introducing chaotic inertia weights and a toroidal walking strategy to balance global search and local exploitation capabilities; the classifier's optimization objective function integrates the basic classification error and periodic consistency constraints, the periodic consistency constraints being achieved by calculating the feature differences of the temporal fusion features in the main period; the classifier takes the temporal fusion features output by the multimodal feature fusion layer as input and outputs the recognition results of the transaction behavior category. The specific implementation is as follows:

[0102] Figure 4 This is a schematic diagram of the structure of the RBF neural network (TW-BA-RBF) optimized by the toroidal walking improved bat algorithm according to an embodiment of the present invention, as shown below. Figure 4 As shown, based on temporal fusion features We construct an RBF neural network classifier optimized by the toroidal walking improved bat algorithm (TW-BA).

[0103] The TW-BA optimized RBF neural network takes as input a fusion of periodic features and transvariable attention. The TW-BA optimization objective is to minimize the classification loss:

[0104]

[0105] The overall classification loss function for transaction behavior recognition is used to optimize the TW-BA algorithm and the RBF neural network, and comprehensively measures the RBF classification error and periodic consistency constraint. The basic classification loss of the RBF neural network (such as cross-entropy loss and mean squared error loss) measures the model's prediction error for transaction behavior categories. : Periodicity constraint coefficient, used to balance the weights of the basic classification loss and the periodicity consistency constraint, with a value range of [value missing]. . Top- The number of main cycles is consistent with the number of main cycles selected in the cycle decoupling process (usually 10). ). Temporal fusion features in the first The feature vectors of each time step contain information on periodic patterns, cross-variable correlations, and long-term time series dependencies. Temporal fusion features in the first Feature vectors at each time step For the first The length of each main cycle is used to reflect the consistency of the cycle. The square of the norm is used to calculate the difference in eigenvectors at different time steps within the same period, measuring periodic consistency.

[0106] Its core updates include:

[0107] Chaotic inertia weights: ;

[0108] Speed ​​updates: ;

[0109] Location update (circular movement): ;

[0110] : No. During the nth iteration, the 1st The optimized individual Chaotic inertial weights of dimensional features are used to explore and develop balancing algorithms. : No. During the nth iteration, the 1st The optimized individual Chaotic inertial weights for dimensional features. Chaos control parameters, satisfying To ensure the chaotic nature of weight updates, the value is fixed at 3.58 here. : Optimize the individual index, corresponding to the individual bat numbers participating in the search in the TW-BA algorithm. Feature dimension index, and Feature Dimensions correspond. : No. During the nth iteration, the 1st The optimized individual A velocity vector of a dimensional feature is used to update the individual's position. : No. During the nth iteration, the 1st The optimized individual The velocity vector of the dimensional feature. : No. During the nth iteration, the 1st The optimized individual The position vector of the dimensional feature corresponds to the parameters to be optimized in the RBF neural network (such as the kernel center and output layer weights). : No. The global optimal position (global optimal solution) of the dimensional feature, i.e., the th feature in all iterations. The optimal parameter values ​​corresponding to the dimensional features, and Dimensions correspond. : No. During the nth iteration, the 1st The optimized individual The frequency of the dimensional feature is used to adjust the step size of the speed update, and its value ranges from 1 to 2. (usually) ). : No. During the nth iteration, the 1st The optimized individual The frequency of the dimensional feature is used to adjust the step size of the speed update, and its value ranges from 1 to 2. (usually) ). : No. During the nth iteration, the 1st The optimized individual The updated position of the dimensional feature is the optimized parameter of the RBF neural network. : No. During the nth iteration, the 1st The loudness of each optimized individual is used to control the magnitude of position updates, and its value ranges from [value range missing]. The loudness decreases as the iteration progresses (reflecting the characteristic that the loudness decreases when the bat approaches the target). : Circular walking control parameters, used to adjust the influence intensity of the circular random numbers, with a value range of [value range missing]. . : Random numbers distributed in a toroidal pattern, following the rules of... The uniform distribution across intervals introduces random exploration characteristics into position updates, ensuring the algorithm's global search capability. During each position update, a new random number is independently generated for each dimension of each individual. The role of T(0,1) is to introduce random perturbation, ensuring that the exploration of the toroidal walk is uncertain and omnidirectional. "Toroidal" refers to its purpose, not its probability distribution itself. This operation is equivalent to introducing a controlled random perturbation in the direction pointing to the current optimal solution. iterBest : No. The optimal position of the feature in the current iteration, i.e., the position of the feature in the current iteration. The optimal parameter values ​​corresponding to the dimensional features.

[0111] The output of the RBF neural network is:

[0112]

[0113] : RBF neural network in the first The output of each time step corresponds to the prediction result of the transaction behavior category (0 - normal, 1 - malicious quoting, 2 - default on performance, 3 - account theft). The number of RBF kernel functions, i.e. the number of hidden layer neurons, is used to fit the nonlinear mapping relationship between transaction behavior features and category labels. The first output layer of the RBF neural network The weight, connected to the first weight The kernel function outputs and the final prediction results. Gaussian kernel function (radial basis function) is used to map high-dimensional temporal fusion features to a low-dimensional feature space. Temporal fusion features in the first The feature vectors at each time step are used as input to the RBF neural network. : No. The kernel center of each RBF kernel function is, in order to... Vectors of the same dimension are determined by optimization using the TW-BA algorithm. : No. The kernel width of each RBF kernel function is used to control the influence range of the kernel function and is determined by optimization using the TW-BA algorithm. With nuclear center of Norm measures the distance between the input feature and the kernel center. The bias term in the output layer of the RBF neural network is used to adjust the baseline of the prediction results and improve the model's fitting ability.

[0114] In this embodiment of the invention, the toroidal walking strategy is a core improvement in position updating, and its specific execution steps are as follows:

[0115] Torus Space Definition: The entire D-dimensional search space of the algorithm is considered as a D-dimensional torus. On this torus, when an individual's position is updated and exceeds the boundary of a certain dimension, it will re-enter the search space from the other side, rather than being truncated by the boundary. For the d-th dimension, its search range is... The period of the torus in this dimension is defined as When the updated position When the boundary is exceeded, modular arithmetic is used to map it to the torus space, correcting it to the search range. The formula is modified as follows:

[0116]

[0117] This modular operation ensures that when a position exceeds the boundary, it can be re-entered from the other side, just like on a torus, which is key to achieving global exploration. For example, if Then, after correction: ;like Then, after correction: .

[0118] Triggered by circular walking: The position update term in the standard bat algorithm ( Above this, multiply by a toroidal distributed random number T(0,1) and the toroidal walking control parameter L. -,Right now .

[0119] This strategy, by simulating "walking" on a toroidal surface, effectively prevents the algorithm from stagnating at the search boundary, forcing the individual to explore the global scope, thus significantly improving the algorithm's ability to escape local optima. The parameter L‾ controls the intensity of this toroidal walking behavior.

[0120] Step 4.2: Credit Assessment Model

[0121] The behavior recognition and credit assessment layer 140 also includes a credit assessment module 142, which contains a credit assessment model for aggregating and mapping temporal fusion features to generate a credit score for the transaction entity. Specifically, it executes the following process: average pooling of the input temporal fusion features across all time steps to obtain a comprehensive feature 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 weighting and normalizing the initial credit score based on a preset scoring indicator weighting system to generate a final credit score ranging from 0 to 100. The scoring indicator weighting system includes at least transaction compliance, fulfillment rate, operational stability, and reliable records. The credit assessment module 142 is implemented through the following credit assessment model:

[0122] Construct a credit scoring system (0-100 points) using time-averaged features from time-series fusion. As input, it is mapped to a credit score through a fully connected layer:

[0123]

[0124]

[0125] The time average vector of the time series fusion features is obtained by averaging the time series fusion features of all time steps, and comprehensively reflects the overall performance of the trading entity within a period. The time step of time series data, i.e. the number of observation points within a single trading period, is consistent with the definition in step three. : For the 1st to the 1st Temporal fusion features at each time step Summation is performed to provide a basis for calculating time averaging. Score: The initial credit score, obtained by mapping from the fully connected layer to a nonlinear transformation, is not normalized and has no explicit limit on its range. The weight matrix of the fully connected layer is used to average the time features. Mapped to an intermediate feature space, with dimension (For intermediate feature dimensions). The bias vector of the fully connected layer is used to adjust the output baseline of the fully connected layer, with dimensions equal to or greater than the intermediate feature dimensions. Consistent. Hyperbolic tangent activation function, range of values It is used to perform nonlinear transformations on the output of fully connected layers, thereby enhancing the model's feature representation capabilities. The scoring mapping weight matrix is ​​used to map the intermediate features after nonlinear transformation to the initial credit score; its dimension is... . The rating mapping bias term is used to adjust the overall offset of the initial credit score, thereby improving the flexibility of the rating mapping.

[0126] The weighting of the scoring indicators is as follows: transaction compliance ( ), fulfillment rate ( ), operational stability ( Trusted Records ).

[0127] Transaction compliance (30%): One of the credit scoring indicators, with a weight of 30%, reflecting the authenticity of the data submitted by the transaction entity and the compliance of the transaction process.

[0128] Fulfillment rate (40%): A core indicator of credit scoring, with a weight of 40%, reflecting the proportion of electricity delivery and payment settlement completed by the trading entity as agreed.

[0129] Operational stability (20%): One of the credit scoring indicators, with a weight of 20%, reflecting the output stability of distributed power generation units, grid connection adaptability, and other conditions.

[0130] Trusted Records (10%): One of the credit scoring indicators, with a weight of 10%, reflects the trusted verification records of the blockchain account of the transaction entity (such as identity authentication, margin payment, node endorsement, etc.).

[0131] The final score is normalized and mapped to the [0, 100] interval:

[0132]

[0133] The final credit score, after normalization, ranges from [value missing]. Credit ratings are divided into the lowest level and 100 points as the highest level. The minimum initial credit score, i.e. the lowest initial score among all transaction entities, is used for the lower limit calibration of normalization. The maximum initial credit score, i.e., the highest initial score among all transaction entities, is used for the upper limit calibration of normalization. Normalization operation linearly maps the initial credit score (Score) to... The scoring range is set to ensure that the scoring range is consistent and comparable.

[0134] Step 4.3: Dynamic Update Mechanism

[0135] The behavior recognition and credit assessment layer 140 also includes a dynamic update module 143, used to jointly update the transaction behavior recognition module 141 and the credit assessment module using newly generated data on a transaction cycle, and to associate the update results with the blockchain account. The specific implementation is as follows:

[0136] Each completed trading cycle The model is automatically updated using the following formula:

[0137]

[0138] : No. The updated model parameter set includes learnable parameters for all modules, such as SSAE, LSTM, RBF neural network, and transvariable attention module. : No. The set of model parameters before the next update, i.e. the current parameter state to be updated. The transaction cycle is the time interval for dynamic updates of the model (such as daily, weekly, or monthly cycles), which is consistent with the actual settlement cycle of distributed power transactions. The learning rate controls the step size for parameter updates; its value typically ranges from [value range missing]. This avoids model oscillations caused by excessively large step sizes or slow convergence caused by excessively small step sizes. Loss function Regarding model parameters The gradient reflects the direction and extent of the impact of parameter changes on the loss, providing guidance for parameter updates. The model update loss function is typically composed of credit assessment loss (such as mean squared error loss) and transaction behavior recognition loss. The weighted sum is used to measure the prediction error of the model in the current period. Training data for the current trading period, including multimodal time-series data, trading behavior labels, and real credit rating labels within that period.

[0139] The update result is linked to the blockchain account of the transaction entity.

[0140] The blockchain storage and verification layer 150 stores preprocessed multimodal time-series data, time-series fusion features, behavior recognition results, and credit scores on the blockchain, and achieves trusted data verification and identity authentication through consensus mechanisms and encryption algorithms. The specific execution process is as follows: Output data from the data preprocessing layer 120, multimodal feature fusion layer 130, and behavior recognition and credit assessment layer 140 are encapsulated into blocks by constructing a Merkle tree structure, and the block header hash value is calculated to ensure data integrity and tamper-proof storage; a proof-of-stake consensus mechanism is adopted, with preset key nodes endorsing and verifying the credit assessment results. Only when the proportion of nodes agreeing to verification reaches or exceeds a preset threshold is the credit assessment result confirmed as valid; identity authentication of transaction entities is achieved through asymmetric encryption algorithms. Transaction entities use their private keys to sign data, and network nodes use the corresponding public keys to verify the validity of the signature; a data interface with external distributed power trading systems is provided, supporting credit score queries, abnormal behavior warnings, and blockchain ledger traceability functions.

[0141] Step 5: Construct the blockchain storage and verification layer 150

[0142] Step 5.1 Data On-Chain Storage: The preprocessed raw data, multimodal fusion features, behavior recognition results, credit scores, and update logs are encapsulated into blocks using a Merkle tree structure. The block header hash is calculated as follows:

[0143]

[0144] Hash The block header hash value is a unique identifier for the current block, calculated using the SHA256 algorithm, and is used to ensure the integrity and immutability of the block data. SHA256 hash algorithm: A cryptographic hash function that takes data of arbitrary length as input and outputs a fixed 256-bit hash value. It is irreversible and collision-resistant. PrevHash: The hash value of the previous block's header, used to connect the current block to the main blockchain, forming a chain structure to ensure data time-series traceability. MerkleRoot: The Merkle root hash value, calculated from the hash values ​​of all transaction-related data (preprocessed raw data, fused features, behavioral recognition results, credit scores, etc.) within the current block using a Merkle tree structure. It is used to quickly verify the integrity of data within the block. Timestamp: A timestamp that records the generation time of the current block (accurate to seconds / milliseconds), used to mark the chronological order of data on-chain, supporting transaction traceability. Nonce: A random number used to meet the computational power verification requirements of the blockchain consensus mechanism (adapted to PoS consensus here, mainly used to distinguish different blocks and avoid hash collisions).

[0145] Distributed storage is implemented using LevelDB and CouchDB to ensure that data is immutable.

[0146] LevelDB: A key-value database used to store blockchain state data (such as account information and parameter configurations), featuring high write performance and low latency.

[0147] CouchDB: A document-oriented database used to store complete block data of a blockchain (such as raw data, complex structured data such as fusion features), supporting complex queries and data backup.

[0148] Step 5.2: Trusted Verification Mechanism

[0149] The PoS consensus mechanism is adopted, and the credit assessment results are endorsed and verified by the transaction center's heavy nodes. The verification pass condition is:

[0150]

[0151] The number of heavy nodes that agree to the verification, i.e., the number of heavy nodes that recognize the validity of the current credit assessment result among the participating verification nodes. The total number of heavy nodes participating in the verification is the total number of heavy nodes in the blockchain network that participate in the verification of the credit assessment results (it must meet the Byzantine fault tolerance requirement, and is usually no less than 5). Consensus threshold, value (Right now (With the agreement of the above nodes), ensuring the consistency and credibility of the verification results, conforming to the Byzantine fault tolerance mechanism of distributed systems. PoS: Proof of Stake, a consensus algorithm for blockchain networks, allocates accounting and verification rights based on the stakes of key nodes (such as node points, deposit amounts), and is more energy-efficient than PoW. Key nodes: Core nodes in the blockchain network that possess complete ledger storage, participate in consensus verification, and transaction clearing calculations, deployed by trusted institutions such as trading centers and key market participants.

[0152] Authentication of transaction participants using asymmetric encryption algorithms: private key signing Data Public key verification .

[0153] Sign: The result of the private key signature of the transaction entity is obtained by the transaction entity using its own private key to encrypt key data (such as transaction declaration data, credit score query request), and is used for identity authentication and non-repudiation. : The private key of the transaction entity, which is a private key for an asymmetric encryption algorithm. It is kept independently by the transaction entity and must not be disclosed. It is used to generate digital signatures. Data: The core data to be signed / verified, including sensitive data requiring identity authentication such as transaction declaration information, credit score results, and account operation requests. : Public key verification function, used to verify the validity of the signature. It takes the public key of the transaction subject, the private key signature, and the original data as input, and outputs a boolean value (True / False). : The public key of the transaction entity, which is the public key of an asymmetric encryption algorithm. It can be publicly distributed in the blockchain network and used to verify the legitimacy of the private key signature. True: Verification passed. When the output is True, it indicates that the signature is valid, the identity of the transaction entity is legitimate, and the data has not been tampered with.

[0154] Step 5.3: Interface Adaptation

[0155] It provides a unified interface with the distributed power trading system, supports functions such as credit score query, abnormal behavior warning, and blockchain ledger traceability, and connects with business processes such as transaction declaration and clearing settlement.

[0156] Unified Interface: The interface between the system and the distributed power trading system (such as a RESTful API) supports standardized data interaction and ensures compatibility.

[0157] Credit Score Inquiry: One of the API functions, allowing transaction entities and transaction centers to query the final credit score of a specified entity via a blockchain address. And rating update log.

[0158] Abnormal Behavior Warning: One of the interface functions, when the behavior recognition result... When an abnormal category is identified (1 - malicious bidding, 2 - breach of contract, 3 - account theft), an early warning message will be automatically sent to the trading center and relevant entities.

[0159] Blockchain ledger traceability: one of the interface functions, supporting traceability via block hash. The system allows users to query complete data within a block using timestamps and other metrics, enabling full traceability of the transaction process.

[0160] After a power purchaser submits their settlement application, the relevant settlement smart contract is invoked, which retrieves the corresponding transaction data from the metering node. After confirmation by the power seller, settlement is processed, and the results are searchable. In case of disputes, the trading center reviews and approves the application. Once confirmed, the settlement is complete and stored on the blockchain, and the agreed amount is automatically transferred to the power seller's account. The power purchaser's process includes the application stage, settlement stage, and dispute resolution stage. During the application stage, the trading center publishes the settlement rules smart contract, the power seller confirms the settlement rules within the smart contract, and the power purchaser invokes the relevant smart contract, metering data, and settlement application. During the settlement stage, the user can query the settlement results, and the trading center publishes the results. In the dispute resolution stage, the trading center determines the dispute outcome based on the settlement data stored on the blockchain. The specific process is as follows: Figure 5 As shown.

[0161] like Figure 6 As shown, each of the 5 nodes sends 10,000 data entries, for a total of 50,000 data entries. Figure 7 As shown, all 50,000 data entries were packaged into the blockchain network, with a data processing success rate of 100%. The largest block size was 1,270,089 bytes, the smallest block size was 130,004 bytes, and the average block size was 804,988.81 bytes. Analysis of block generation time reveals that each block is generated approximately every 5 seconds, indicating a consensus performance of 5 seconds per block. Therefore, the maximum data processing performance is 1436.6 tps, and the average data processing performance is 909.09 tps.

[0162] The distributed power transaction behavior recognition and credit assessment system based on multimodal neural networks provided in this invention utilizes an improved SSAE-LSTM architecture. Through intramodal sparse regularization and cross-modal commonality extraction, it achieves deeper multimodal fusion and more accurate feature representation, effectively solving the problem of multimodal data heterogeneity. Combined with LSTM's ability to capture temporal dependencies, feature fusion retains the unique characteristics of each modality while uncovering deep cross-modal correlations, providing high-quality feature support for behavior recognition and credit assessment. Compared to traditional fusion methods, the effectiveness of feature representation is significantly improved. The cross-variable attention enhancement module (CvAM), guided by the trusted verification modality, dynamically allocates weights for transaction behavior and operational state modalities, accurately adapting the dynamic coupling relationship of the three types of variables. It achieves high accuracy in identifying abnormal behaviors such as malicious bidding and default, with more precise dynamic correlation capture and more sensitive anomaly identification. Compared to static feature fusion methods, the anomaly identification latency is reduced to within 1 minute, and sensitivity is improved. The TW-BA algorithm, through chaotic inertial weights and a toroidal walking strategy, achieves more efficient classification optimization and stronger generalization ability. It effectively balances global exploration and local exploitation capabilities, preventing RBF neural networks from getting trapped in local optima. Compared to standard BA and PSO optimization algorithms, it reduces classification loss and significantly improves generalization ability across different trading scenarios, adapting to the complex and ever-changing scenarios of distributed power trading. The credit assessment system is deeply integrated with real-time trading behavior, reflecting the dynamic changes in the credit status of entities through a periodic update mechanism. Combined with blockchain technology, it ensures data storage, identity authentication, and the immutability of assessment results, guaranteeing the fairness and transparency of credit scoring. It also supports intelligent applications such as automatic binding of trading permissions and automatic handling of defaults, providing full-process technical support for risk management in the distributed power trading market.

[0163] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A distributed power trading behavior recognition and credit assessment system based on multimodal neural networks, characterized in that, include: The data acquisition layer is used to collect multimodal time-series data of the entire distributed power trading chain in real time. The multimodal time-series data includes: trading behavior modal data, operating status modal data, and trusted verification modal data. The data preprocessing layer is used to perform data cleaning, standardization, and timestamp-based time-series alignment on the multimodal time-series data to generate multimodal sample data with uniform scale and synchronized time series. The multimodal feature fusion layer adopts an improved SSAE-LSTM fusion architecture for multimodal sample data to achieve feature extraction within each modality and cross-modal feature fusion, so as to extract the periodic patterns, cross-modal dynamic correlations and long-term dependencies contained therein, and output temporal fusion features. The behavior recognition and credit assessment layer, based on the aforementioned temporal fusion features, uses a radial basis function neural network optimized by the toroidal walking improved bat algorithm to classify and identify transaction behaviors, outputs behavior recognition results, and calculates credit scores through a dynamic credit assessment model. The blockchain storage and verification layer is used to store preprocessed multimodal time-series data, time-series fusion features, behavior recognition results, and credit scores on the blockchain, and to achieve trusted data verification and identity authentication through consensus mechanisms and encryption algorithms.

2. The system according to claim 1, characterized in that, The multimodal feature fusion layer specifically includes: The intramodal feature extraction module uses the first layer of a stacked sparse autoencoder to independently train three types of modal data: transaction behavior, running status, and trusted verification, and extracts the specific feature vectors of each modality. The cross-modal feature fusion module concatenates the specific feature vectors of the three modalities and inputs them into the second layer of the stacked sparse autoencoder (SSAE) to extract cross-modal shared features. The periodic decoupling and filtering module performs periodic decoupling on the cross-modal shared features through fast Fourier transform to filter the main periodic features and reconstructs them into a three-dimensional tensor, embedding a learnable periodic index code. The cross-variable attention enhancement module uses trusted verification modal features as queries, transaction behavior and runtime status modal features as keys and values, dynamically calculates attention weights, and outputs enhanced cross-variable fusion features. The temporal dependency modeling module uses a long short-term memory network (LSTM) to perform temporal modeling on the cross-variable fusion features, captures the long-term temporal dependencies after fusion, and outputs the temporal fusion features.

3. The system according to claim 2, characterized in that, The intra-modal feature extraction module specifically performs the following process: Independent stacked sparse autoencoders were used for training three modalities: transaction behavior, operational status, and trusted verification. The training of each stacked sparse autoencoder aims to minimize the comprehensive loss function, which includes at least the reconstruction error of the input data, the regularization penalty term of the model weights, and the sparsity constraint on the activation values ​​of the hidden layer neurons. Through training, the original time-series data of each modality is encoded into a modality-specific feature vector, which serves as the input to the cross-modality feature fusion module.

4. The system according to claim 2, characterized in that, The periodic decoupling and filtering module is configured to perform frequency domain transformation and principal period analysis on the cross-modal shared feature representation, specifically including: The cross-modal shared feature representation is transformed from the time domain to the frequency domain using a fast Fourier transform, and the average amplitude intensity of each frequency component in different feature dimensions is calculated. A preset number of core main cycles are selected based on the average amplitude intensity; Based on each selected master period, the shared feature representation is reorganized to construct a three-dimensional tensor containing feature dimensions, number of period segments, and period length. Learnable periodic index encodings are embedded in each periodic segment of the three-dimensional tensor to distinguish and identify different periodic patterns.

5. The system according to claim 2, characterized in that, The cross-variable attention enhancement module specifically performs the following process: Use a subset of shared features from trusted verification modalities as the attention query vector; The shared feature subsets of the transaction behavior mode and the running state mode are used together as the attention key vector and value vector; By calculating the similarity between the query vector and the key vector, the weights of transaction behavior and running status features in the value vector are dynamically allocated, thereby achieving feature enhancement fusion guided by credible verification information and generating enhanced cross-variable fusion features, which serve as input to the time-series dependency modeling module.

6. The system according to claim 2, characterized in that, The time-series dependency modeling module specifically executes the following process: The cross-variable fusion features are input into a multi-layer long short-term memory network; The gating mechanism in the multi-layer long short-term memory network enables the filtering, updating, and transmission of historical information to capture long-range dependencies in time-series data. The hidden states of all time steps are spliced ​​together to form a comprehensive time series fusion feature that includes periodic patterns, cross-variable correlations, and long-term dependencies.

7. The system according to claim 1, characterized in that, The behavior recognition and credit assessment layer includes: The transaction behavior recognition module uses a radial basis function neural network optimized by a toroidal walking improved bat algorithm to classify and recognize transaction behavior, and outputs the behavior recognition results. The credit assessment module includes a credit assessment model, which is used to aggregate and map the time-series fusion features to generate a credit score for the transaction entity. The dynamic update module is used to jointly update the transaction behavior identification module and the credit assessment module according to the transaction cycle using newly generated data, and associate the update results with the blockchain account.

8. The system according to claim 7, characterized in that, The transaction behavior identification module specifically performs the following process: A radial basis function neural network optimized with an improved bat algorithm is used as the classifier; The improved bat algorithm updates and optimizes the individual's position by introducing chaotic inertial weights and a toroidal walking strategy, in order to balance global search and local exploitation capabilities. The optimization objective function of the classifier combines the basic classification error and the periodic consistency constraint, which is achieved by calculating the feature difference of the temporal fusion features in the main period. The classifier takes the temporal fusion features output by the multimodal feature fusion layer as input and outputs the identification result of the transaction behavior category.

9. The system according to claim 7, characterized in that... The credit assessment module executes the following process: The input temporal fusion features are averaged over all time steps to obtain a comprehensive feature vector; The comprehensive feature vector is input into a fully connected neural network for nonlinear transformation; Map the output of the fully connected network to the initial credit score; Based on a preset scoring indicator weighting system, the initial credit score is weighted and normalized to generate a final credit score ranging from 0 to 100; wherein the scoring indicator weighting system includes at least transaction compliance, fulfillment rate, operational stability and reliable records.

10. The system according to claim 1, characterized in that, The blockchain storage and verification layer is configured to perform the following process: The output data from the data preprocessing layer, the multimodal feature fusion layer, and the behavior recognition and credit assessment layer are encapsulated into blocks by constructing a Merkle tree structure, and the block header hash value is calculated to achieve data integrity and tamper-proof storage. The credit assessment result is confirmed as valid only when the proportion of nodes agreeing to the verification reaches or exceeds a preset threshold. The identity authentication of the transaction subjects is achieved through asymmetric encryption algorithms. The transaction subjects use their private keys to sign the data, and the network nodes use the corresponding public keys to verify the validity of the signature. It provides a data interface with external distributed power trading systems, and supports credit score query, abnormal behavior warning and blockchain ledger traceability functions.

Images