Intelligent auxiliary decision-making system and method for power transaction based on multivariate data

Abstract

The invention discloses a multivariate data-based intelligent auxiliary decision-making system and a multivariate data-based intelligent auxiliary decision-making method for power transaction, relates to the technical field of power transaction, and aims to solve the technical problem of inaccurate prediction caused by the fact that a current traditional model cannot effectively process multi-time scale dependence and spatial relevance. The intelligent prediction and pre-decision module is used for realizing electricity price prediction and auxiliary decision-making functions, and the intelligent prediction and pre-decision module comprises an electricity price prediction model module and an auxiliary decision-making application module. According to the method, the time dependence learning module adopts a multi-sampling mechanism and a hierarchical contrast learning strategy, multi-granularity, multi-scale and down-sampling extracted time features are fused, the spatial dependence learning module is based on an implicit graph structure adaptive construction method, and diffusion convolution and adaptive adjacency matrix dynamic modeling are combined, so that the time-dependent learning module is used for learning time features of the multi-granularity, multi-scale and down-sampling extracted time features. The problem that prediction is inaccurate due to the fact that a current traditional model cannot effectively process multi-time-scale dependence and spatial relevance is solved.

Description

Technical Field

[0001] This invention relates to the field of power trading technology, and more specifically, to a power trading intelligent auxiliary decision-making system and method based on multi-source data. Background Technology

[0002] Guided by the global energy transition and the "dual carbon" goal, electricity market reform has become a core path to optimize resource allocation and promote the consumption of renewable energy. With the increasing number of electricity market participants, the diversification of transaction types, and the trend towards shorter and faster transaction cycles, the complexity and uncertainty of electricity trading have significantly increased, and electricity price fluctuations have continued to widen, posing severe challenges to market participants' decision-making and risk management.

[0003] The industry commonly employs statistical time series analysis methods (such as the ARIMA model) and basic neural network structures for electricity price forecasting. However, electricity market price fluctuations are simultaneously influenced by multiple time scales, including ultra-short-term real-time trading, short-term day-ahead trading, and medium- to long-term monthly or annual trading. Traditional single-scale time series models cannot effectively capture this complex multi-granularity time dependence, resulting in weak predictive power for price inflection points and abnormal fluctuations. Moreover, the electricity market is essentially a complex network system with strong spatial correlations; the electricity price at a given node is not only affected by local supply and demand but also closely related to the power flow, transmission constraints, and trading behavior of neighboring nodes. Most existing models treat each node as an independent entity, ignoring the spatial dependence caused by the power grid topology, leading to predictions that deviate from actual physical laws. Therefore, we propose an intelligent auxiliary decision-making system and method for electricity trading based on multivariate data. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology, adapt to the needs of reality, and provide a power trading intelligent auxiliary decision-making system and method based on multi-source data, so as to solve the technical problem that the current traditional model cannot effectively handle multi-time scale dependence and spatial correlation, resulting in inaccurate prediction.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an intelligent auxiliary decision-making system for power trading based on multivariate data. This system includes an intelligent prediction and pre-decision-making module, which is used to implement electricity price prediction and auxiliary decision-making functions. The intelligent prediction and pre-decision-making module includes an electricity price prediction model module and an auxiliary decision-making application module. The electricity price prediction model module improves prediction accuracy through multi-sampling, graph structure learning, and pre-training techniques. The electricity price prediction model module includes a time-dependency learning module, a spatial dependency learning module, and a sequence enhancement representation module. The time-dependency learning module employs a multi-sampling mechanism unit in conjunction with a feature fusion unit, utilizing multi-granularity sampling, multi-scale sampling, and... The system extracts temporal features using a combination of downsampling and pooling layers to capture temporal correlations and improve the accuracy of multivariate time series prediction. The spatial dependency learning module, based on an implicit graph structure adaptive construction unit and a node similarity measurement unit, learns node features and inter-node similarity metrics to construct a dynamic graph structure. The sequence enhancement representation module, based on a Transformer pre-training unit, constructs a pre-trained model through a Transformer autoencoder and employs a masked autoencoder strategy unit to extract high-order features of the time series. The auxiliary decision application module integrates an electricity price prediction model to provide real-time decision support to users, including electricity monitoring, risk assessment, and transaction strategy optimization functions.

[0006] Preferably, the spatial dependency learning module uses a graph convolutional neural network to capture spatial dependencies, employing a diffusing convolution formula: ,in The signal transfer matrix is ​​shown in the figure. The node feature matrix, For the first Step convolution weights, by randomly initializing node embeddings , The formula is generated by combining the activation function and normalization operation: ;

[0007] Convolutional layers: .

[0008] Preferably, the time-dependent learning module employs hierarchical contrastive learning, and the loss function includes time-contrast loss and instance-contrast loss;

[0009] The time-comparison loss is expressed as:

[0010]

[0011] The instance contrast loss is expressed as:

[0012]

[0013] The overall loss function is:

[0014]

[0015] Among them, time comparison loss is... The index of the input time series sample. For timestamps.

[0016] Preferably, it also includes a data layer, which adopts a hybrid storage strategy to store multi-source heterogeneous raw data, standardized data after governance, and model prediction results. The data layer includes a multi-source spatiotemporal data aggregation and governance module and a distributed storage module. The multi-source spatiotemporal data aggregation and governance module is used to collect and process multi-source heterogeneous data, including historical electricity trading prices, location information of new energy power plants, regional meteorological spatiotemporal data, and electricity supply and demand ratio data. It constructs a data source collection pool through knowledge ontology theory and uses reinforcement learning algorithms to optimize the data source priority.

[0017] Preferably, the multi-dimensional spatiotemporal data aggregation and governance module includes a data source acquisition pool construction module, a spatiotemporal pattern association identification module, and a time series high-precision smoothing module. The data source acquisition pool construction module constructs a data source acquisition pool based on a knowledge ontology mapping unit and uses a reinforcement learning optimization unit to dynamically optimize the data source priority and association relationship. The spatiotemporal pattern association identification module constructs an implicit relationship automatic identification model through an implicit relationship automatic identification unit and uses spatiotemporal transformer technology. It also constructs a multi-level spatiotemporal dependency capture model by combining a multi-level dependency capture unit perceptual network. The time series high-precision smoothing module is based on a noise removal unit and an adaptive resampling unit. It uses orthogonal polynomial projection technology to realize discrete data continuity, high-frequency noise removal, and adaptive resampling operations.

[0018] Preferably, the data source acquisition pool construction module uses web crawling technology to capture multi-source data, stores time-series data in an InfluxDB database, stores unstructured and semi-structured data in a NoSQL database, and uses the Q-learning algorithm to iteratively optimize the data source.

[0019] Preferably, the auxiliary decision-making application module includes a real-time monitoring module, a market analysis module, a risk assessment module, and a transaction management module. The real-time monitoring module senses power consumption data and equipment status monitoring data through IoT devices based on a power consumption sensing unit and an equipment status monitoring unit. The market analysis module collects market fundamental information through distributed crawler technology based on a fundamental information collection unit and an automated price analysis unit. The prediction decision-making module simulates transactions and quantitatively analyzes risks through an electricity price prediction model output unit and a risk assessment unit. The transaction management module provides candlestick chart visualization and replay functions.

[0020] Preferably, it also includes a service layer, which is used to transform data resources into intelligent services that meet the needs of power trading. The service layer includes an intelligent prediction and decision-making module and a security authentication module. The security authentication module uses multiple technical means to prevent risks such as data leakage and unauthorized access.

[0021] Preferably, it also includes a communication layer and a presentation layer. The communication layer enables efficient data transmission between modules and between the system and the outside world, and the presentation layer enables visual functionality between modules and the system through visualization function modules.

[0022] A smart auxiliary decision-making method for power trading based on multivariate data, comprising the following steps:

[0023] S100, Data Aggregation and Governance Phase: By achieving standardized integration of multi-source heterogeneous data, high-quality input is provided for subsequent analysis;

[0024] S101. Data Acquisition and Priority Optimization: A data source acquisition pool is constructed based on the knowledge ontology mapping theory. Multi-dimensional spatiotemporal data is acquired in real time through distributed crawler technology. Reinforcement learning algorithms are used to dynamically optimize the data source priority to ensure the timeliness and accuracy of data acquisition. Time-series data is stored in the InfluxDB database, and unstructured data is stored in the NoSQL database.

[0025] S102, Spatiotemporal Relationship Mining: Construct an automatic identification model for implicit relationships using spatiotemporal transformer technology, mine spatiotemporal dependencies between data through self-attention mechanism, establish a multi-level spatiotemporal dependency capture model, and improve the depth of data correlation analysis;

[0026] S103: Time series optimization processing: The Legendre polynomial projection technique is applied to realize the continuity of discrete data, the high-frequency noise removal algorithm is used to improve data quality, and the adaptive resampling technique is used to ensure data consistency.

[0027] S200, Intelligent Prediction Model Construction Process: Achieve accurate electricity price prediction through learning involving time, space and sequence features.

[0028] S201, Time-dependent learning: Employs a multi-sampling mechanism including multi-granularity, multi-scale, and downsampling, combined with a hierarchical contrastive learning strategy, to capture the long-term and periodic features of time series.

[0029] S202, Spatial Dependency Learning: Based on the implicit graph structure adaptive construction method, it uses graph convolutional networks to learn the spatial dependencies between variables and dynamically models node associations through an adaptive adjacency matrix;

[0030] S203. Sequence Augmentation Representation: Pre-training is performed using a Transformer autoencoder, and a masked autoencoder strategy is used to extract high-order features of the time series to enhance the model's generalization ability.

[0031] S300, Transform the forecast results into practical decision support;

[0032] S301, Real-time Monitoring and Market Analysis: Integrates IoT devices to collect real-time power data, combines distributed crawler technology to obtain market fundamentals information, and generates dynamic analysis reports;

[0033] S302, Risk Assessment and Simulated Trading: Based on the prediction results, assess trading risks through quantitative analysis models (such as Value at Risk calculation), and provide simulated trading and replay functions;

[0034] S303, Visualization and Interaction Support: Utilize visualization tools such as candlestick charts to display electricity price trends and risk assessment results, assisting users in optimizing their strategies.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. This invention employs a multi-sampling mechanism and hierarchical contrastive learning strategy in its time dependency learning module, integrating time features extracted from multiple granularities, scales, and downsampling. This fully explores the long-term trends, periodic features, and trend characteristics of time series, significantly enhancing the ability to represent time dependencies. The spatial dependency learning module, based on an implicit graph structure adaptive construction method, combines diffusing convolution and adaptive adjacency matrix dynamic modeling to accurately capture the spatial dependencies between multivariate variables, solving the problem that traditional explicit graph structures are ill-suited to dynamic changes. The sequence enhancement representation module, pre-trained using Transformer autoencoders and mask autoencoders, extracts high-order semantic features of time series, effectively improving the model's generalization ability and prediction accuracy. This meets the precise prediction needs of actual transactions and solves the problem that current traditional models cannot effectively handle multi-timescale dependencies and spatial correlations, leading to inaccurate predictions.

[0037] 2. This invention also integrates accurate forecast results through an auxiliary decision-making application module, combined with real-time monitoring, market analysis, risk assessment, and transaction management functions, providing users with intelligent decision support throughout the entire process; through IoT devices and distributed crawler technology, it realizes real-time collection and dynamic analysis of electricity data and market fundamentals information, helping users to keep abreast of market dynamics; based on quantitative analysis models and simulated trading and replay functions, it effectively assesses the potential risks and returns of trading strategies, helping users optimize trading strategies and reduce market risks; visual tools such as candlestick charts intuitively display electricity price trends and risk assessment results, improving the convenience and intuitiveness of decision-making.

[0038] 3. This invention also achieves adaptive optimization and efficient collection of data sources by combining knowledge ontology theory with Q-learning algorithm, ensuring the timeliness and accuracy of multi-source heterogeneous data, providing a high-quality data foundation for subsequent prediction modeling. The application of spatiotemporal transformer technology and self-attention mechanism can automatically mine the implicit spatiotemporal correlations between data, capturing complex dependencies without manual intervention, improving the depth and efficiency of data correlation analysis. The use of Legendre multinomial projection technology effectively solves problems such as noise interference and inconsistent sampling frequency in sequence data, realizing standardized data processing and providing consistency guarantee for model input. Attached Figure Description

[0039] Figure 1 This is a system block diagram of the present invention;

[0040] Figure 2 This is a detailed system block diagram of the multi-dimensional spatiotemporal data aggregation module of the present invention;

[0041] Figure 3 This is a detailed system block diagram of the electricity price prediction module of the present invention;

[0042] Figure 4 This is a specific system block diagram of the decision support application module of the present invention;

[0043] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0044] Example 1:

[0045] like Figures 1 to 4 As shown, this embodiment provides an intelligent auxiliary decision-making system for power trading based on multivariate data. The system includes an intelligent prediction and pre-decision-making module, which is used to realize electricity price prediction and auxiliary decision-making functions. The intelligent prediction and pre-decision-making module includes an electricity price prediction model module and an auxiliary decision-making application module. The electricity price prediction model module improves prediction accuracy through multi-sampling, graph structure learning, and pre-training techniques. The electricity price prediction model module includes a time dependency learning module, a spatial dependency learning module, and a sequence enhancement representation module. The time dependency learning module uses a multi-sampling mechanism unit in conjunction with a feature fusion unit to extract time features through a combination of multi-granularity sampling, multi-scale sampling, and downsampling, and introduces a pooling layer to capture time correlations to improve the accuracy of multivariate time series prediction. The spatial dependency learning module is based on an implicit graph structure adaptive construction unit and a node similarity measurement unit, learning node features and node similarity measurements to construct a dynamic graph structure.

[0046] In the implicit graph structure adaptive building block, a graph convolutional neural network is used to capture spatial dependencies. Diffusion convolution is employed, which represents the diffusion process of the graph signal as a finite... The step transition matrix, its basic formula is expressed as:

[0047]

[0048] in The signal transfer matrix is ​​shown in the figure. The node feature matrix, For the first Step convolution weights;

[0049] When constructing an undirected graph, In a directed graph, the forward transition matrix is ​​used. and backward transition matrix The formula for diffuse convolution is expressed as:

[0050]

[0051] By randomly initializing node embeddings , The formula is generated by combining the activation function and normalization operation:

[0052]

[0053] pass and Multiply them to obtain the spatial dependency weight between the source node and the target node. Then use... () Activate the function to eliminate weak connections, application The `()` function is used to normalize the adaptive adjacency matrix. The normalized adaptive adjacency matrix is ​​considered as the transformation matrix of the hidden diffusion process. By combining predefined spatial dependencies and self-learned hidden graph dependencies, the following graph convolutional layer is proposed:

[0054]

[0055] If the graph structure is not predefined, the adaptive part of graph convolution can be used directly to capture spatial dependencies.

[0056] The final design incorporates the objective function of the model, which simultaneously considers node sequence information, inter-node relationship information, and graph structure information in the multivariate time series prediction model. After passing through multiple Transformers, it is connected to a fully connected layer to achieve the output nonlinear prediction result.

[0057] By aggregating different time frequencies of multivariate time series, time series at different scales can be obtained. To enable the encoder to better learn the representation of the time series at different scales, hierarchical contrastive learning and a contrastive loss function are used. The loss function under hierarchical contrastive learning consists of two parts: temporal contrastive loss and instance contrastive loss.

[0058] The time comparison loss is to make The index of the input time series sample. For timestamps.

[0059] Timestamp First The time-comparison loss of a time series can be expressed as:

[0060]

[0061] The instance comparison shows that positive sample pairs are close in distance, while negative sample pairs are far apart. Negative sample pairs are generated using random negative sampling within a batch.

[0062]

[0063] in, B is the timestamp, and B is the batch size. Let be the sample set. This formula demonstrates the effectiveness of the multisampling mechanism in enhancing time-dependent diversity and improving prediction accuracy.

[0064] The final loss function is obtained by fusing the two loss functions mentioned above. The entire loss is calculated using a hierarchical contrastive learning approach, with calculation performed after each layer of fusion. .

[0065] The sequence enhancement representation module is based on the Transformer pre-training unit and constructs a pre-trained model through the Transformer autoencoder. It uses a mask autoencoder strategy unit to extract high-order features of time series data. Through self-supervised learning, the model automatically learns the deep-order features of the sequence data. Then, the high-order features are input into the downstream application model to optimize tasks such as meta-time series prediction and classification.

[0066] This invention employs a multi-sampling mechanism and hierarchical contrastive learning strategy in its time dependency learning module, integrating time features extracted from multiple granularities, scales, and downsampling. This fully explores the long-term trends, periodic features, and trend characteristics of time series, significantly enhancing the ability to represent time dependencies. The spatial dependency learning module, based on an implicit graph structure adaptive construction method, combines diffusing convolution and adaptive adjacency matrix dynamic modeling to accurately capture the spatial dependencies between multivariate variables, solving the problem that traditional explicit graph structures cannot adapt to dynamic changes. The sequence enhancement representation module, pre-trained using Transformer autoencoders and mask autoencoders, extracts high-order semantic features of time series, effectively improving the model's generalization ability and prediction accuracy. This meets the accurate prediction needs of actual transactions and solves the problem that current traditional models cannot effectively handle multi-time-scale dependencies and spatial correlations, leading to inaccurate predictions.

[0067] The auxiliary decision-making application module integrates an electricity price forecasting model to provide real-time decision support for users, including electricity monitoring, risk assessment, and trading strategy optimization functions. The auxiliary decision-making application module includes a real-time monitoring module, a market analysis module, a risk assessment module, and a trading management module. The real-time monitoring module uses an electricity sensing unit and an equipment status monitoring unit to sense electricity data and equipment status monitoring data through IoT devices. The market analysis module uses a fundamental information collection unit and an automated price analysis unit to collect market fundamental information through distributed crawler technology. The forecasting decision-making module uses an electricity price forecasting model output unit and a risk assessment unit to simulate trading and quantitatively analyze and assess risks through the electricity price forecasting model. The trading management module provides candlestick chart visualization and replay functions.

[0068] Preferably, it also includes a data layer, which adopts a hybrid storage strategy to store multi-source heterogeneous raw data, standardized data after governance, and model prediction results. The data layer includes a multi-source spatiotemporal data aggregation and governance module and a distributed storage module. The multi-source spatiotemporal data aggregation and governance module is used to collect and process multi-source heterogeneous data, including historical electricity trading prices, new energy power plant location information, regional meteorological spatiotemporal data, and electricity supply and demand ratio data. It constructs a data source collection pool based on knowledge ontology theory and uses reinforcement learning algorithms to optimize data source priorities. The multi-source spatiotemporal data aggregation and governance module includes a data source collection pool construction module, a spatiotemporal pattern association identification module, and a time series high-precision smoothing module. The data source collection pool construction module constructs a data source collection pool based on a knowledge ontology mapping unit and uses a reinforcement learning optimization unit to dynamically optimize data source priorities and association relationships.

[0069] A mapping relationship is established between core concepts in the power trading field (such as electricity prices, meteorological data, and power supply-demand ratio) and multi-source data (historical electricity prices, locations of new energy power plants, regional meteorological spatiotemporal data, etc.); then, multi-source data is crawled using web crawling technology, and stored according to data type. Time-series data is stored in the InfluxDB database, and unstructured and semi-structured data are stored in the NoSQL database.

[0070] The data source acquisition pool construction module uses web crawling technology to collect data from multiple sources, storing time-series data in an InfluxDB database and unstructured and semi-structured data in a NoSQL database. To dynamically optimize data source priority, the Q-learning algorithm from reinforcement learning is introduced, achieving adaptive adjustment through a state-action value function, the formula of which is:

[0071]

[0072] in, Represents the real-time status of the data source. Actions to adjust data source priority, For instant rewards based on data quality, As a discount factor, this formula can demonstrate the adaptive adjustment capability of the data source, effectively improving the efficiency and quality of data collection.

[0073] The spatiotemporal pattern association recognition module constructs an automatic recognition model of implicit relationships through an implicit relationship automatic recognition unit and a spatiotemporal transformer technology, and constructs a multi-level spatiotemporal dependency capture model by combining a multi-level dependency capture unit perception network.

[0074] The time series high-precision smoothing module is based on a noise removal unit and an adaptive resampling unit, and uses orthogonal polynomial projection technology to realize discrete data continuity, high-frequency noise removal and adaptive resampling operations.

[0075] Based on the properties of polynomial projection, changes in high-frequency sequences often require higher-order polynomials for fitting. Since noise typically manifests as high-frequency components, excessively high-order polynomials may also incorporate this noise into the curve during fitting, thus reducing the overall model's generalization ability. Therefore, adjusting the polynomial order can effectively prevent high-frequency noise from being included in the projection curve, thereby filtering and removing sequence noise. In practical applications, finding a suitable polynomial order that ensures effective fitting of the original sequence while minimizing noise removal can be achieved by balancing denoising effectiveness and fitting accuracy using the following error formula to reach the optimal denoising effect and model performance:

[0076]

[0077] in, This indicates the order of the polynomial.

[0078] The orthogonal polynomials are Legendre polynomials. These polynomials reproject the original sampling sequence into a continuous function space. This projection method effectively transforms discrete sampling points into a continuous function representation, facilitating subsequent processing and analysis. Next, utilizing the orthogonality property of Legendre polynomials, the dynamic characteristics of the mapping coefficients are derived. These coefficients reflect the variation of the sampling sequence in the function space.

[0079] The mapping process is described by state-space equations:

[0080]

[0081] in, For mapping coefficients, , For the dynamic matrix, The original discrete sequence is used. High-frequency noise is filtered by adjusting the polynomial order, and the function is optimized. The resampling interval is determined dynamically. Due to sampling bias, This represents the target deviation.

[0082] An automatic implicit relationship identification model is built using spatiotemporal transformer technology, with spatiotemporal data sets as input. , For timestamps, For spatial data, a self-attention mechanism is used to capture the hidden spatiotemporal dependencies between data points, as shown in the formula:

[0083]

[0084] in, (Query matrix) (Key matrix) The (value matrix) is generated by transforming the input data. This is a dimensionality scaling factor to avoid excessively large calculated values. This formula demonstrates the module's ability to uncover implicit relationships in multivariate data without manual intervention.

[0085] By combining knowledge ontology theory with the Q-learning algorithm, adaptive optimization and efficient data collection of data sources are achieved, ensuring the timeliness and accuracy of multi-source heterogeneous data and providing a high-quality data foundation for subsequent predictive modeling. The application of spatiotemporal transformer technology and self-attention mechanism can automatically mine the implicit spatiotemporal correlations between data, capturing complex dependencies without manual intervention, and improving the depth and efficiency of data correlation analysis. The use of Legendre multinomial projection technology effectively solves problems such as noise interference and inconsistent sampling frequencies in sequential data, realizes standardized data processing, and provides consistency assurance for model input.

[0086] Preferably, it also includes a service layer, which is used to transform data resources into intelligent services that meet the needs of power trading. The service layer includes an intelligent prediction and decision-making module and a security authentication module. The security authentication module uses multiple technical means to prevent risks such as data leakage and unauthorized access. It also includes a communication layer and a presentation layer. The communication layer realizes efficient data transmission between modules and between the system and the outside world. The presentation layer realizes the visualization function between modules and the system through a visualization function module.

[0087] Using distributed web crawling technology, the system collects various information related to the electricity market in real time from the internet. This information includes policy changes, economic conditions, market supply and demand fluctuations, power facility operation status, and weather conditions. The system then organizes and analyzes this information to provide comprehensive data support for electricity price forecasting. Users can use this information to better understand the fundamental factors behind the market, thereby making more informed trading decisions.

[0088] Utilizing big data processing technologies such as Apache Spark and combined with advanced machine learning algorithms, this module performs real-time processing and analysis of electricity market price data. It can identify price fluctuation patterns, trends, and anomalies, generating detailed market analysis reports, including price trends, market supply and demand, and analysis of factors influencing prices. These analytical results provide strong decision support for market participants, helping them better understand market dynamics and formulate reasonable electricity purchase and sales strategies.

[0089] Electricity price forecasting model:

[0090] By utilizing advanced machine learning algorithms such as time series analysis, neural networks, and deep learning, a precise electricity price prediction model is established by training on historical electricity price data. This model comprehensively considers multiple factors, such as historical prices, market supply and demand, weather conditions, seasonal variations, and policy impacts, to provide users with scientific predictions of future electricity prices. These predictions help users formulate electricity procurement and sales strategies in advance, reducing market risks.

[0091] Electricity price candlestick chart:

[0092] Utilizing visual programming technology and combined with backend data processing results, this module displays electricity price trends in an intuitive candlestick chart format. It not only shows current and historical price data but also includes information such as price changes, percentage fluctuations, and trading volume over 24 hours. Users can intuitively understand market trends through the candlestick chart, analyze the reasons for price fluctuations, and provide a basis for trading decisions.

[0093] Simulated trading risk assessment and trade replay functions:

[0094] Leveraging big data storage and processing capabilities, the system uses algorithmic models to simulate and evaluate users' trading strategies. It can simulate different market scenarios and assess the potential risks and returns of trading strategies. The replay function allows users to review and analyze past trading decisions, understand the actual performance of their strategies, and learn from and optimize them. These features help users conduct thorough risk assessments and strategy validation before real trading. By integrating these functions, a power price forecasting and decision support system is built, creating a more comprehensive, real-time, and intelligent power market monitoring and decision support platform. This helps enterprises make accurate and efficient decisions in complex power trading environments, promotes the maturity and optimization of regional power markets, and facilitates the effective utilization of renewable energy.

[0095] By integrating accurate forecast results into the decision support application module, and combining real-time monitoring, market analysis, risk assessment, and transaction management functions, it provides users with intelligent decision support throughout the entire process. Through IoT devices and distributed crawler technology, it achieves real-time collection and dynamic analysis of electricity data and market fundamentals, helping users to keep abreast of market dynamics. Based on quantitative analysis models and simulated trading and replay functions, it effectively assesses the potential risks and returns of trading strategies, helping users optimize trading strategies and reduce market risks. Visual tools such as candlestick charts intuitively display electricity price trends and risk assessment results, improving the convenience and intuitiveness of decision-making.

[0096] Example 2:

[0097] like Figure 5 As shown in the figure, this embodiment provides a smart auxiliary decision-making method for power trading based on multi-source data. The method includes the following steps:

[0098] S100, Data Aggregation and Governance Stage: By achieving standardized integration of multi-source heterogeneous data, high-quality input is provided for subsequent analysis.

[0099] S101. Data Acquisition and Priority Optimization: A data source acquisition pool is constructed based on the knowledge ontology mapping theory. Multi-dimensional spatiotemporal data is acquired in real time through distributed crawler technology. Reinforcement learning algorithms are used to dynamically optimize the data source priority to ensure the timeliness and accuracy of data acquisition. Time-series data is stored in the InfluxDB database, and unstructured data is stored in the NoSQL database.

[0100] S102, Spatiotemporal Relationship Mining: Construct an automatic identification model for implicit relationships using spatiotemporal transformer technology, mine spatiotemporal dependencies between data through self-attention mechanism, establish a multi-level spatiotemporal dependency capture model, and improve the depth of data correlation analysis;

[0101] S103: Time series optimization processing: The Legendre polynomial projection technique is applied to realize the continuity of discrete data, the high-frequency noise removal algorithm is used to improve data quality, and the adaptive resampling technique is used to ensure data consistency.

[0102] S200, Intelligent Prediction Model Construction Process: Achieve accurate electricity price prediction through learning involving time, space and sequence features.

[0103] S201, Time-dependent learning: Employs a multi-sampling mechanism including multi-granularity, multi-scale, and downsampling, combined with a hierarchical contrastive learning strategy, to capture the long-term and periodic features of time series.

[0104] S202, Spatial Dependency Learning: Based on the implicit graph structure adaptive construction method, it uses graph convolutional networks to learn the spatial dependencies between variables and dynamically models node associations through an adaptive adjacency matrix;

[0105] S203. Sequence Augmentation Representation: Pre-training is performed using a Transformer autoencoder, and a masked autoencoder strategy is used to extract high-order features of the time series to enhance the model's generalization ability.

[0106] S300, Transform the forecast results into practical decision support;

[0107] S301, Real-time Monitoring and Market Analysis: Integrates IoT devices to collect real-time power data, combines distributed crawler technology to obtain market fundamentals information, and generates dynamic analysis reports;

[0108] S302, Risk Assessment and Simulated Trading: Based on the prediction results, assess trading risks through quantitative analysis models (such as Value at Risk calculation), and provide simulated trading and replay functions;

[0109] S303, Visualization and Interaction Support: Utilize visualization tools such as candlestick charts to display electricity price trends and risk assessment results, assisting users in optimizing their strategies.

[0110] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A smart auxiliary decision-making system for power trading based on multi-source data, characterized in that, The system includes an intelligent forecasting and pre-decision-making module, which is used to realize electricity price forecasting and decision support functions; the intelligent forecasting and pre-decision-making module includes an electricity price forecasting model module and a decision support application module; The electricity price prediction model module improves prediction accuracy through multi-sampling, graph structure learning, and pre-training techniques. The electricity price prediction model module includes a time-dependent learning module, a spatial-dependent learning module, and a sequence augmentation representation module; The time-dependent learning module employs a multi-sampling mechanism unit in conjunction with a feature fusion unit. It extracts time features through a combination of multi-granularity sampling, multi-scale sampling, and downsampling, and introduces a pooling layer to capture time correlations in order to improve the accuracy of multivariate time series prediction. The spatial dependency learning module learns node features and the similarity measure between nodes based on the implicit graph structure adaptive construction unit and the node similarity measurement unit, and constructs a dynamic graph structure. The sequence enhancement representation module is based on the Transformer pre-training unit and constructs a pre-trained model through the Transformer autoencoder. It uses a mask autoencoder strategy unit to extract high-order features of time series. The auxiliary decision-making application module integrates an electricity price forecasting model to provide real-time decision support for users, including electricity monitoring, risk assessment, and trading strategy optimization functions.

2. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 1, characterized in that, The spatial dependency learning module uses a graph convolutional neural network to capture spatial dependencies, employing a diffusing convolution formula: ,in The signal transfer matrix is ​​shown in the figure. The node feature matrix, For the first Step convolution weights; By randomly initializing node embeddings , The formula is generated by combining the activation function and normalization operation: ; Image convolutional layer: 。 3. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 2, characterized in that, The time-dependent learning module employs hierarchical contrastive learning, and its loss function includes time-contrast loss and instance-contrast loss. The time-comparison loss is expressed as: The instance-contrast loss is expressed as: The overall loss function is: Among them, time comparison loss is... The index of the input time series sample. For timestamps.

4. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 3, characterized in that, It also includes a data layer, which adopts a hybrid storage strategy to store multi-source heterogeneous raw data, standardized data after governance, and model prediction results; The data layer includes a multi-dimensional spatiotemporal data aggregation and governance module and a distributed storage module; The multi-source spatiotemporal data aggregation and governance module is used to collect and process multi-source heterogeneous data, including historical electricity prices, location information of new energy power plants, regional meteorological spatiotemporal data, and electricity supply and demand ratio data. It constructs a data source collection pool through knowledge ontology theory and uses reinforcement learning algorithms to optimize the data source priority.

5. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 4, characterized in that, The multi-dimensional spatiotemporal data aggregation and governance module includes a data source acquisition pool construction module, a spatiotemporal pattern association identification module, and a time series high-precision smoothing module; The data source acquisition pool construction module constructs a data source acquisition pool based on a knowledge ontology mapping unit, and uses a reinforcement learning optimization unit to dynamically optimize the data source priority and association relationship. The spatiotemporal pattern association recognition module constructs an automatic recognition model of implicit relationships through an implicit relationship automatic recognition unit and a spatiotemporal transformer technology, and constructs a multi-level spatiotemporal dependency capture model by combining a multi-level dependency capture unit perception network. The time series high-precision smoothing module is based on a noise removal unit and an adaptive resampling unit, and uses orthogonal polynomial projection technology to realize discrete data continuity, high-frequency noise removal and adaptive resampling operations.

6. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 5, characterized in that, The data source acquisition pool construction module uses web crawling technology to capture multi-source data, stores time-series data in an InfluxDB database, stores unstructured and semi-structured data in a NoSQL database, and uses the Q-learning algorithm to iteratively optimize the data source.

7. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 6, characterized in that, The decision support application module includes a real-time monitoring module, a market analysis module, a risk assessment module, and a transaction management module; The real-time monitoring module senses power data and device status monitoring data through IoT devices based on the power sensing unit and the device status monitoring unit. The market analysis module, based on the fundamental information collection unit and the automated price analysis unit, collects market fundamental information through distributed crawler technology. The prediction and decision-making module, based on the electricity price prediction model output unit and the risk assessment unit, simulates transactions and quantitatively analyzes and assesses risks through the electricity price prediction model. The transaction management module provides candlestick chart visualization and replay functions.

8. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 7, characterized in that, It also includes a service layer, which is used to transform data resources into intelligent services that meet the needs of power trading. The service layer includes an intelligent prediction and decision-making module and a security authentication module. The security authentication module uses multiple technical means to prevent risks such as data leakage and unauthorized access.

9. The intelligent auxiliary decision-making system for power trading based on multi-source data according to claim 8, characterized in that, It also includes a communication layer and a presentation layer; The communication layer enables efficient data transmission between modules and between the system and the outside world; The presentation layer enables visualization between modules and the system through visualization function modules.

10. A power trading intelligent auxiliary decision-making method based on multi-source data, applicable to the power trading intelligent auxiliary decision-making system based on multi-source data as described in any one of claims 1-9, characterized in that, The method includes the following steps: S100, Data Aggregation and Governance Phase: By achieving standardized integration of multi-source heterogeneous data, high-quality input is provided for subsequent analysis; S101. Data Acquisition and Priority Optimization: A data source acquisition pool is constructed based on the knowledge ontology mapping theory. Multi-dimensional spatiotemporal data is acquired in real time through distributed crawler technology. Reinforcement learning algorithms are used to dynamically optimize the data source priority to ensure the timeliness and accuracy of data acquisition. Time-series data is stored in the InfluxDB database, and unstructured data is stored in the NoSQL database. S102, Spatiotemporal Relationship Mining: Construct an automatic identification model for implicit relationships using spatiotemporal transformer technology, mine spatiotemporal dependencies between data through self-attention mechanism, establish a multi-level spatiotemporal dependency capture model, and improve the depth of data correlation analysis; S103: Time series optimization processing: The Legendre polynomial projection technique is applied to realize the continuity of discrete data, the high-frequency noise removal algorithm is used to improve data quality, and the adaptive resampling technique is used to ensure data consistency. S200, Intelligent Prediction Model Construction Process: Achieve accurate electricity price prediction through learning involving time, space and sequence features; S201, Time-dependent learning: Employs a multi-sampling mechanism including multi-granularity, multi-scale, and downsampling, combined with a hierarchical contrastive learning strategy, to capture the long-term and periodic features of time series. S202, Spatial Dependency Learning: Based on the implicit graph structure adaptive construction method, it uses graph convolutional networks to learn the spatial dependencies between variables and dynamically models node associations through an adaptive adjacency matrix; S203. Sequence Augmentation Representation: Pre-training is performed using a Transformer autoencoder, and a masked autoencoder strategy is used to extract high-order features of the time series to enhance the model's generalization ability. S300, Transform the forecast results into practical decision support; S301, Real-time Monitoring and Market Analysis: Integrates IoT devices to collect real-time power data, combines distributed crawler technology to obtain market fundamentals information, and generates dynamic analysis reports; S302. Risk Assessment and Simulated Trading: Based on the prediction results, the trading risk is assessed through a quantitative analysis model, and simulated trading and replay functions are provided. S303, Visualization and Interaction Support: Utilize visualization tools such as candlestick charts to display electricity price trends and risk assessment results, assisting users in optimizing their strategies.

Images