Power transaction auxiliary decision system and method based on multi-source data and intelligent algorithm

Abstract

The application discloses a power transaction auxiliary decision system and method based on multi-source data and intelligent algorithm, and belongs to the field of power systems.The power transaction auxiliary decision system and method based on multi-source data and intelligent algorithm comprises a data fusion module, an intelligent prediction module, a transaction execution module and a knowledge management module.The application solves the problems of low power transaction decision efficiency, weak risk control and difficult maximization of benefits caused by insufficient information processing capacity, single and isolated prediction model, disconnection between decision and execution and lack of continuous learning mechanism in the prior art, and realizes the collaborative prediction of market core elements, the automatic generation and execution of risk adaptive strategies and the self-iteration optimization of the system based on practical feedback by constructing an integrated auxiliary decision system of data driving, intelligent prediction, automatic execution and closed-loop learning, so that precise, efficient and stable transaction decision support is provided for power market participants.

Description

Technical Field

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

[0002] Against the backdrop of deepening power market reform, the trading instruments in the power spot market are becoming increasingly diverse, trading frequency is significantly accelerating, and price volatility and uncertainty are greatly increasing. Traditional power trading decisions mainly rely on manual experience and simple statistical analysis, making it difficult to effectively integrate and process multi-source heterogeneous information such as market clearing data, grid topology and operational constraints, high-precision weather forecasts, real-time generator status, and massive historical trading data. Existing technical solutions mostly focus on single-point forecasting of load or electricity price, lacking the ability to coordinate and forecast core market elements (such as load, renewable energy output, and nodal prices), and the forecast results are difficult to directly translate into executable and traceable trading strategies. In addition, market participants face challenges such as complex rules, short decision windows, and numerous risk factors. Manual operation is inefficient and prone to errors, making it difficult to achieve a refined balance between returns and risks; therefore, it cannot meet the current needs. To address this, we propose a power trading auxiliary decision-making system and method based on multi-source data and intelligent algorithms. Summary of the Invention

[0003] The purpose of this invention is to provide a power trading auxiliary decision-making system and method based on multi-source data and intelligent algorithms. By constructing a collaborative closed loop of four core modules—data fusion, intelligent prediction, transaction execution, and knowledge management—it achieves efficient integration of multi-source heterogeneous market data, coupled and collaborative prediction of core market elements, automatic optimization and execution of risk-adaptive trading strategies, and continuous learning and optimization of the decision-making process, thus solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a power trading auxiliary decision-making system based on multi-source data and intelligent algorithms, comprising:

[0005] The data fusion module is configured to collect multi-source data from electricity spot market operators, meteorological departments, and enterprise internal management systems, and to clean, align, and timestamp the collected multi-source data, generating unique hash values ​​for data entries and storing them in a distributed time-series database. The multi-source data includes at least market clearing data, grid boundary conditions, weather forecast data, generator unit operating parameters, and historical transaction data.

[0006] The intelligent forecasting module is configured to access multi-source data processed by the data fusion module. It constructs a coupled forecasting model based on deep learning and reinforcement learning to simultaneously forecast the unified dispatch load, new energy output, and nodal marginal electricity price for a specific future period. It also integrates a market rule reasoning engine based on knowledge graphs and a strategy optimizer based on multi-objective particle swarm optimization algorithm to generate and compare multiple alternative day-ahead and real-time trading quotation strategies based on the forecast results, user-defined risk preferences, and cost constraints.

[0007] The transaction execution module is configured to interface with the power trading platform API, receive the preferred strategies output by the intelligent prediction module, convert the strategy instructions into executable standard messages, and automatically execute order placement, order cancellation and transaction confirmation operations when preset market conditions are met through the embedded smart contract engine, and track market transaction status and position status in real time.

[0008] The knowledge management module is configured to perform automatic review and analysis on a daily, time-period, and transaction basis. By comparing the actual clearing results with the prediction strategy, it quantitatively evaluates the performance of the strategy and stores the review conclusions in the knowledge graph in association with key market scenarios for iterative optimization of the prediction model and strategy rules.

[0009] Furthermore, the data fusion module includes:

[0010] The data acquisition module is deployed on the power plant's edge computing node and is configured to collect and preprocess unit operation data and station meteorological data in real time, and upload them to the cloud data center through an encrypted channel.

[0011] The quality verification module is configured to use a data anomaly detection model based on statistical rules and machine learning to verify the integrity, rationality and consistency of collected multi-source data, and to repair missing and abnormal multi-source data using a spatiotemporal correlation interpolation algorithm.

[0012] Furthermore, the intelligent prediction module includes:

[0013] The coupled prediction module is configured to build the coupled prediction model based on a multi-task learning framework. It uses a shared feature extraction layer to encode the spatiotemporal correlation features of multi-source data and outputs the probability distribution prediction results of future cycles simultaneously through load prediction branch, new energy output prediction branch and electricity price prediction branch respectively.

[0014] The strategy optimization module is configured to receive probability distribution prediction results and combine them with real-time market rule constraints and user risk preferences obtained from the knowledge graph inference engine to construct an optimization problem with the goal of maximizing expected returns and minimizing risk value. The multi-objective particle swarm optimization algorithm is used to solve the problem and output a set of day-ahead and real-time candidate pricing strategies.

[0015] Furthermore, a shared feature extraction layer is used to encode the spatiotemporal correlation features of multi-source data, and the probability distribution prediction results for future cycles are simultaneously output through load forecasting, renewable energy output forecasting, and electricity price forecasting branches, respectively, including:

[0016] Calculate the mutual information value and time lag correlation coefficient between each data source in the multi-source data. When the mutual information value exceeds the first threshold, it is determined that there is a dependency relationship. The time delay is determined according to the peak position of the time lag correlation coefficient. Construct a directed spatiotemporal dependency graph, where nodes represent data sources, the weight of the edges is the normalized mutual information value, and the direction of the edges is determined by the time lag order.

[0017] Multi-source data are weighted and fused according to the edge weights in the directed spatiotemporal dependency graph. The fusion weight is proportional to the edge weight. Time alignment compensation is performed on data source pairs whose edge weights exceed the second threshold to generate fused feature representations.

[0018] The fused feature representation is input into the shared feature extraction layer to generate task-specific weighted feature vectors for three tasks: load forecasting, renewable energy output forecasting, and electricity price forecasting, as well as a task-shared general feature vector.

[0019] Based on the supply and demand balance of the power system and the electricity price formation mechanism, the constraints are transformed into differentiable penalty terms. According to the specific weighted feature vectors of each task and the common feature vectors shared by the tasks, the load forecasting branch, the new energy output forecasting branch, and the electricity price forecasting branch with cross-constraint penalty terms are determined.

[0020] The load probability distribution parameters are determined through the load forecasting branch; the renewable energy output probability distribution parameters are determined through the renewable energy output forecasting branch, using the load probability distribution parameters as additional inputs; and the electricity price probability distribution parameters are determined through the electricity price forecasting branch, using both the load probability distribution parameters and the renewable energy output probability parameters as additional inputs.

[0021] A preset number of future scenarios are generated by sampling from three probability distribution parameters. The net load value is calculated for each scenario. The net load value and the sampled electricity price value are checked to see if they satisfy a monotonically increasing relationship. The sampled value of new energy output is also checked to see if it is within the physical constraints. Inconsistency indicators are statistically analyzed.

[0022] When the inconsistency index exceeds the third threshold, adjust the probability distribution parameters or truncate the distribution tails that exceed the physical constraints, and repeat the physical consistency check until the inconsistency index drops below the third threshold or reaches the preset number of iterations, and determine the corrected probability distribution parameters.

[0023] The corrected probability distribution parameters are aligned along the time axis, and the confidence intervals and covariance matrices at each time point are calculated. A structured prediction dataset containing the predicted mean, confidence intervals, covariance matrix, and inconsistency index is constructed as the probability distribution prediction result for future periods.

[0024] Furthermore, the strategy optimization module executes the following process:

[0025] The probability distribution prediction results output by the coupled prediction module are transmitted to the strategy optimization module as uncertainty quantification input.

[0026] The strategy optimization module generates multiple future market scenarios that conform to the distribution based on probability distribution prediction results through simulation.

[0027] In each simulation scenario, a stochastic optimization model is constructed by combining scenario-based constraint rules obtained in real time from the knowledge graph inference engine and the user's preset risk preferences, with the goal of maximizing expected utility.

[0028] A multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, and a set of day-to-day and real-time trading strategies with robustness under different risk preferences is output.

[0029] Furthermore, a multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, outputting a set of day-to-day and real-time trading strategies that are robust under different risk preferences, including:

[0030] Based on the constraint boundary of the decision variables in the stochastic optimization model, random sampling is performed in the multidimensional policy space to initialize the initial position and initial velocity of each particle in the particle swarm algorithm, thereby obtaining the initial particle state vector set.

[0031] Substitute the position vector of each particle in the initial particle state vector set into the stochastic optimization model for simulation evaluation, calculate the objective function value of maximizing expected return and the objective function value of risk at each state position, and determine the multi-objective fitness evaluation matrix that includes the two-dimensional objective value associated with each particle.

[0032] Calculate the dominance relationships among the multi-objective fitness evaluation matrix, update the matrix based on the comparison of dominance relationships, and determine the individual historical optimal position and global optimal position of each particle.

[0033] Based on the individual historical best position and the global best position, the position migration state of each particle after a set number of iterations in the multi-dimensional strategy space is continuously calculated to determine the Pareto optimal solution set.

[0034] Based on the Pareto optimal solution set and the pre-introduced user multi-level risk preference coefficients, the target position vector is determined by the linear weighted matching method, and the target position vector is mapped to a set of day-to-day and real-time trading strategies that are robust under different risk preferences.

[0035] Furthermore, the transaction execution module includes:

[0036] The strategy compilation module is configured to compile the received preferred strategy parameters into standardized trading instruction messages that conform to the communication protocol of the power trading platform;

[0037] The contract execution module is configured to have dynamically configurable smart contract scripts deployed internally. The smart contract scripts encapsulate complete transaction logic for order placement, cancellation, and confirmation, and can monitor and judge multiple conditions such as preset price, time, trading volume, and price difference of related products in parallel based on real-time access to market data streams.

[0038] The condition triggering module is configured to automatically invoke standardized trading instruction messages when market data is detected to meet the triggering conditions of any smart contract, execute the corresponding trading operation through the power trading platform API, and simultaneously generate an operation log with timestamps and execution environment hash values.

[0039] Furthermore, the knowledge management module includes:

[0040] The post-trade analysis module is configured to automatically perform the following analyses at three levels: daily, time-segment, and single transaction, based on transaction execution logs and actual market clearing data:

[0041] Attribution analysis is performed on the deviation between the predicted values ​​and the actual values ​​output by the coupled prediction model;

[0042] Compare the expected returns of the trading strategy with the actual risk-adjusted returns and attribute the performance.

[0043] The knowledge graph update module is configured to structure and map the attribution conclusions, key performance indicators and corresponding market scenario characteristics obtained from the retrospective analysis module into entity nodes, attribute relationships and confidence labels in the knowledge graph.

[0044] Among them, when the confidence level of the relationship between a specific scenario and a strategy pattern reaches a preset threshold, a model optimization instruction is triggered;

[0045] Update the driver module to respond to model optimization commands, convert high-confidence knowledge graph triple data into an incremental learning sample set for coupled prediction models and constraint rules for policy optimization, and drive targeted iterative updates of coupled prediction models and policy rules.

[0046] A power trading auxiliary decision-making method based on multi-source data and intelligent algorithms is applied to a power trading auxiliary decision-making system based on multi-source data and intelligent algorithms. The method includes the following steps:

[0047] Market, meteorological, and unit operation data are collected through a cloud-edge collaborative architecture, and after cleaning, verification, time alignment, and hash identification, they are stored in a distributed storage layer.

[0048] The trained coupled prediction model is invoked to make synchronous predictions of future market factors. Combined with market rules and risk constraints inferred from the knowledge graph, a multi-objective optimization algorithm is used to generate multiple candidate trading strategies and conduct simulation comparisons to output a recommended strategy set.

[0049] The recommended strategy is transformed into a smart contract script with trigger conditions, which is then deployed in the transaction execution module. When market conditions are met, the order placement operation is automatically executed, and the execution status and transaction results are monitored throughout the process.

[0050] Based on actual transaction data and market results, the system performs automated retrospective analysis on prediction accuracy, strategy effectiveness, and execution compliance. The analysis results are then structured and fed back to a knowledge graph and coupled prediction model to enable the system to learn and continuously optimize itself.

[0051] Furthermore, the trained coupled prediction model is invoked to simultaneously predict future market factors, specifically including:

[0052] Based on a multi-task learning framework, a shared feature layer is used to extract spatiotemporal correlation features from multi-source data.

[0053] The probability distributions of load, renewable energy output, and electricity price are predicted by outputting them through dedicated network branches.

[0054] The forecast results and uncertainty range are integrated with current policy constraints, unit physical limitations, and contractual obligations retrieved from the knowledge graph;

[0055] By taking the maximization of expected returns and the control of risk value as optimization objectives, a constrained swarm intelligence algorithm is used to search the policy space and output the optimal policy frontier for decision-makers to choose from.

[0056] Furthermore, the entire execution status and transaction results are monitored, specifically including:

[0057] Parameterize the pricing curve, power consumption, and triggering logic in the strategy;

[0058] Generate smart contract code and deploy it in a trusted execution environment, with the contract state connected to real-time market data streams;

[0059] When market data is detected to meet the contract trigger conditions, the contract automatically calls the trading platform API to send compliant order placement instructions;

[0060] Record the entire lifecycle log of the contract from creation and triggering to completion, and calculate the hash value for evidence storage.

[0061] Furthermore, based on actual transaction data and market results, automated post-mortem analysis is conducted to assess forecast accuracy, strategy effectiveness, and execution compliance, specifically including:

[0062] Deviation analysis is performed between predicted and actual values ​​to identify sources of systematic error and adjust model features and structure.

[0063] Compare the simulated returns of the strategy with the risk-adjusted returns after actual implementation to analyze whether the performance difference stems from forecast bias or unexpected market events.

[0064] The new market patterns, effective strategy models, and lessons learned during this trading cycle will be constructed into new entities, relationships, and rules in a knowledge graph, and their confidence levels will be labeled.

[0065] When the confidence level of new knowledge exceeds a set threshold, the incremental learning process of the coupled prediction model and the rule base update of the policy optimizer are triggered.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] This invention utilizes multi-source data fusion technology to effectively integrate multi-dimensional heterogeneous information from the market, meteorology, power grid, and generating units, thereby overcoming the problems of data silos and low processing efficiency in traditional methods. Through a coupled prediction model based on multi-task learning and reinforcement learning, it synchronously and accurately predicts core market elements such as load, renewable energy output, and nodal electricity prices. This solves the shortcomings of existing technologies where single-point prediction ignores the correlation between elements, leading to large strategy biases. Furthermore, by combining knowledge graphs and multi-objective optimization algorithms, the prediction results can be transformed into multiple executable trading strategies that consider user risk preferences and market constraints, thus achieving a seamless transition from prediction to strategy. This improves the scientific nature and operability of decision-making. Smart contracts enable automatic execution of strategies and automated review based on actual trading results, feeding back experiential knowledge to the system for continuous self-optimization. This significantly improves the decision-making efficiency, return stability, and risk response capabilities of market participants. Attached Figure Description

[0068] Figure 1 This is a structural diagram of the power trading auxiliary decision-making system based on multi-source data and intelligent algorithms of the present invention;

[0069] Figure 2 The flowchart shows the power trading auxiliary decision-making method based on multi-source data and intelligent algorithms of the present invention.

[0070] Figure 3 This is an execution diagram of the power trading auxiliary decision-making method based on multi-source data and intelligent algorithms of the present invention. Detailed Implementation

[0071] 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.

[0072] To address the technical problems in existing power trading technologies, such as insufficient information processing capabilities, simplistic and isolated predictive models, disconnect between decision-making and execution, and a lack of continuous learning mechanisms, which result in low efficiency in power trading decisions, weak risk control, and difficulty in maximizing returns, please refer to [link to relevant documentation]. Figures 1-3 This embodiment provides the following technical solution:

[0073] A power trading auxiliary decision-making system based on multi-source data and intelligent algorithms includes:

[0074] The data fusion module is configured to collect multi-source data from electricity spot market operators, meteorological departments, and enterprise internal management systems, and to clean, align, and timestamp the collected multi-source data, generating unique hash values ​​for data entries and storing them in a distributed time-series database. The multi-source data includes at least market clearing data, grid boundary conditions, weather forecast data, generator unit operating parameters, and historical transaction data.

[0075] Specifically, for raw data streams collected in real-time or near real-time from different data sources, preliminary cleaning is performed according to predefined integrity, rationality, and consistency verification rules to remove records with obvious errors, abnormal formats, and those exceeding the physical limits. A timestamp synchronization mechanism based on a unified reference clock is used to align all data entries to a unified millisecond-level time granularity. For time sequence misalignments caused by transmission delays, interpolation correction is performed using the spatiotemporal correlation of data from adjacent time periods. For each cleaned and time-aligned data entry, its key content fields (including timestamp, data type, value, and data source identifier) ​​are extracted and concatenated into a string of a specific format. The SHA-256 hash algorithm is used to calculate a unique and irreversible hash value for this string, which serves as the globally unique identifier and data integrity verification credential for this data entry in the distributed time-series database.

[0076] The intelligent forecasting module is configured to access multi-source data processed by the data fusion module. Based on a coupled forecasting model built using deep learning and reinforcement learning, it is used to simultaneously forecast the unified dispatch load, renewable energy output, and nodal marginal electricity price for a specific future period. It also integrates a market rule reasoning engine based on a knowledge graph and a strategy optimizer based on a multi-objective particle swarm optimization algorithm. Based on the forecast results, user-defined risk preferences, and cost constraints, it generates and compares multiple alternative day-ahead and real-time trading quotation strategies.

[0077] The transaction execution module is configured to interface with the power trading platform API, receive the preferred strategies output by the intelligent prediction module, convert the strategy instructions into executable standard messages, and automatically execute order placement, order cancellation and transaction confirmation operations when preset market conditions are met through the embedded smart contract engine, and track market transaction status and position status in real time.

[0078] The knowledge management module is configured to perform automatic review and analysis on a daily, time-period, and transaction basis. By comparing the actual clearing results with the prediction strategy, it quantitatively evaluates the performance of the strategy and stores the review conclusions in the knowledge graph in association with key market scenarios for iterative optimization of the prediction model and strategy rules.

[0079] The technical effects of the above-mentioned technical solution are as follows: the data fusion module enables unified cleaning, alignment, and reliable storage of multi-source heterogeneous data, thereby providing a high-quality data foundation for decision-making; the intelligent prediction module, with the help of coupled prediction models and rule inference engines, enables high-precision synchronous prediction of key market indicators and optimization of multi-objective strategies, thereby improving the adaptability and robustness of pricing strategies; the transaction execution module realizes automated, standardized execution and real-time monitoring of strategies through APIs and smart contracts, thereby reducing the risk of manual operation and improving response efficiency; and the knowledge management module drives the closed-loop iterative optimization of coupled prediction models and trading strategies through continuous review and scenario-based knowledge accumulation, thereby comprehensively improving the overall intelligence level, risk control capability, and long-term adaptability of power trading decisions.

[0080] The data fusion module includes:

[0081] The data acquisition module is deployed on the power plant's edge computing node and is configured to collect and preprocess unit operation data and station meteorological data in real time, and upload them to the cloud data center through an encrypted channel.

[0082] The quality verification module is configured to use a data anomaly detection model based on statistical rules and machine learning to verify the completeness, rationality, and consistency of collected multi-source data, and to repair missing and abnormal multi-source data using a spatiotemporal correlation interpolation algorithm. Specifically:

[0083] The data stream is checked for completeness, rationality and consistency according to the preset statistical rules. The completeness check is achieved by checking the matching degree between the number of data points and the sampling frequency. The rationality check is achieved by judging whether the data value is within the physical or historical experience threshold range. The consistency check is achieved by comparing whether the numerical deviation of the same indicator from different related data sources at the same time exceeds the threshold.

[0084] Meanwhile, a machine learning-based data anomaly detection model runs in parallel. This data anomaly detection model is trained using historical normal data. After receiving the data stream in real time, it extracts time-series features, identifies whether data points deviate from the normal pattern through an unsupervised learning algorithm, and cross-validates with the statistical rule detection results to confirm the anomaly.

[0085] For data points marked as missing or anomalous, a spatiotemporal correlation interpolation algorithm is activated. Based on the geographical location, device type, and normal data from nearby time points of the problematic data points, a spatiotemporal weight matrix is ​​constructed. By weighted fusion of valid observations from similar devices in adjacent time periods and from adjacent devices in the same time period, the most reasonable repair value is generated for filling, ensuring the continuity and reliability of the data stream.

[0086] The technical effects of the above solution are as follows: the data acquisition module deployed on the power plant side realizes localized real-time preprocessing and secure transmission of unit and meteorological data, thereby reducing the cloud processing load and ensuring data timeliness and source security. The quality verification module uses a detection model that integrates rules and intelligent algorithms to perform in-depth verification and repair of multi-source data, which can improve the integrity, rationality and consistency of data, thus providing a reliable high-quality data foundation for intelligent analysis.

[0087] The intelligent prediction module includes:

[0088] The coupled prediction module is configured to build the coupled prediction model based on a multi-task learning framework. It uses a shared feature extraction layer to encode the spatiotemporal correlation features of multi-source data and simultaneously outputs the probability distribution prediction results for future cycles through load prediction, renewable energy output prediction, and electricity price prediction branches, respectively. Specifically:

[0089] Deploy a coupled prediction model based on a multi-task learning framework. The input of the coupled prediction model is a spatiotemporally aligned multi-source data sequence from the data fusion module.

[0090] The coupled prediction model uses a shared spatiotemporal feature extraction layer (which may include spatiotemporal graph convolutional networks or attention mechanisms) for unified processing to deeply encode and fuse the complex spatiotemporal coupling relationships and interactions between heterogeneous data such as market clearing data, power grid topology, weather forecasts, historical loads, historical output of new energy sources and electricity prices.

[0091] The coupled prediction model feeds the learned shared high-dimensional feature representations into three structurally independent dedicated prediction branch networks (i.e., load prediction branch, renewable energy output prediction branch, and electricity price prediction branch). Each branch is designed to output the probability distribution of the target variable for a specific future prediction period (e.g., 96 points in the next 24 hours). This is achieved by outputting quantile values ​​or parameterized distributions (e.g., the mean and variance of a Gaussian distribution), thereby simultaneously generating prediction results with uncertainty quantification for the centrally dispatched load, renewable energy output, and nodal marginal electricity price.

[0092] The strategy optimization module is configured to receive probability distribution prediction results and, in conjunction with real-time market rule constraints and user risk preferences obtained from the knowledge graph inference engine, construct an optimization problem with the objective of maximizing expected return and minimizing risk value. The problem is solved using a multi-objective particle swarm optimization algorithm, outputting a set of day-ahead and real-time candidate pricing strategies for comparison and selection. The specific process is as follows:

[0093] The probability distribution prediction results output by the coupled prediction module are transmitted to the strategy optimization module as uncertainty quantification input.

[0094] The strategy optimization module generates multiple future market scenarios that conform to the distribution based on probability distribution prediction results through simulation.

[0095] In each simulation scenario, a stochastic optimization model is constructed by combining scenario-based constraint rules obtained in real time from the knowledge graph inference engine and the user's preset risk preferences, with the goal of maximizing expected utility.

[0096] A multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, and a set of day-to-day and real-time trading strategies with robustness under different risk preferences is output.

[0097] The technical effects of the above solution are as follows: By using the coupled prediction module to synchronously generate probability distribution predictions of load, new energy output, and electricity price based on a multi-task learning framework, it can capture the inherent correlation and uncertainty among core market elements, thereby providing more comprehensive forward-looking information for decision-making. The strategy optimization module uses probability distribution to simulate multiple scenarios and combines real-time rules and user preferences to build a stochastic optimization model. Through a multi-objective particle swarm optimization algorithm, it solves a series of robust strategies that balance returns and risks, thereby realizing a closed-loop generation from uncertainty prediction to optimal decision-making strategy, thus improving the scientific nature, adaptability, and risk resistance of the trading strategy.

[0098] Specifically, a shared feature extraction layer is used to encode the spatiotemporal correlation features of multi-source data, and the probability distribution prediction results for future cycles are simultaneously output through load forecasting, renewable energy output forecasting, and electricity price forecasting branches, including:

[0099] Calculate the mutual information value and time lag correlation coefficient between each data source in the multi-source data. When the mutual information value exceeds the first threshold, it is determined that there is a dependency relationship. The time delay is determined according to the peak position of the time lag correlation coefficient. Construct a directed spatiotemporal dependency graph, where nodes represent data sources, the weight of the edges is the normalized mutual information value, and the direction of the edges is determined by the time lag order.

[0100] The multi-source data includes historical load sequences, historical renewable energy output sequences, historical electricity price sequences, meteorological forecast data (temperature, humidity, wind speed, irradiance, etc.), date and time characteristics (weekdays, holiday markers, etc.), and unit operating status data. All data is collected at an hourly granularity, spanning the past 90 days. First, each data source is standardized to normalize the numerical range to a unified interval. Then, the mutual information values ​​and time-lag correlation coefficients between the data sources are calculated.

[0101] The mutual information value between data sources is calculated using a nearest neighbor estimation method, reflecting the statistical dependence between the two data sources. For each pair of data sources, the correlation coefficient is calculated at different time lags, ranging from 0 to 24 hours. By analyzing the correlation coefficient curves over time lags, the temporal relationship between the data sources can be identified. A first threshold is set; when the mutual information value exceeds this threshold, a significant dependency between the two data sources is determined. In this embodiment, the first threshold is set to 0.3. Among the data source pairs with a dependency, the peak position of the time-lag correlation coefficient is found; the time lag corresponding to this position is the time delay between the data sources. For example, if the correlation coefficient between temperature and load reaches its peak at a lag of 2 hours, then temperature is determined to lead load by 2 hours.

[0102] Construct a directed graph where nodes represent data sources and edges represent data source pairs with dependencies. The weight of each edge is defined as a normalized mutual information value, achieved by dividing each mutual information value by the maximum mutual information value. The direction of the edges is determined by their time lag order: the earlier data source points to the later data source.

[0103] Multi-source data are weighted and fused based on the edge weights in the directed spatiotemporal dependency graph. The fusion weight is proportional to the edge weight. For data source pairs whose edge weights exceed the second threshold, time alignment compensation is performed to generate fused feature representations.

[0104] For each data source, collect all edges pointing to it and calculate the fusion weight based on the edge weights. The fusion weight is proportional to the edge weights and is normalized so that the sum of all fusion weights pointing to the same data source is 1.

[0105] A second threshold is set, and time alignment compensation is performed on data source pairs whose edge weights exceed this threshold. In this embodiment, the second threshold is set to 0.6. Specifically, the leading data source is shifted backward by the corresponding number of time steps based on the time delay, so that it is aligned with the lagging data source in time. This alignment operation can eliminate the time lag effect, allowing the fused features to better reflect the synchronization relationship between the data sources.

[0106] For each data source, a weighted sum is applied to all data sources pointing to it (after time alignment compensation) to generate a fused feature. The fused features from all data sources are then concatenated into a fused feature matrix.

[0107] The fused feature representation is input into the shared feature extraction layer to generate task-specific weighted feature vectors for three tasks: load forecasting, renewable energy output forecasting, and electricity price forecasting, as well as a task-shared general feature vector.

[0108] The fused feature matrix is ​​input into a shared feature extraction layer, which employs an attention mechanism to achieve task-oriented feature weighting. A multi-layer neural network then performs a non-linear transformation on the fused features to extract high-level semantic features. This embodiment uses a Transformer architecture, comprising multiple encoder layers, each containing a multi-head attention mechanism and a feedforward neural network.

[0109] For the three tasks of load forecasting, renewable energy output forecasting, and electricity price forecasting, learnable query vectors are defined for each task. An attention mechanism is used to calculate the attention score for each task on the feature dimension; this score reflects the importance of different features to each task.

[0110] For each task, the attention scores are used to weight and sum the features, resulting in three sets of task-specific weighted feature vectors. This mechanism allows each task to adaptively select the features that are most important to it.

[0111] The extracted high-level features are subjected to global pooling to obtain a task-shared common feature vector, which contains common information of all tasks.

[0112] Based on the supply and demand balance of the power system and the electricity price formation mechanism, the constraints are transformed into differentiable penalty terms. According to the specific weighted feature vectors of each task and the common feature vectors shared by the tasks, the load forecasting branch, the renewable energy output forecasting branch, and the electricity price forecasting branch with cross-constraint penalty terms are determined.

[0113] At any given time, the load under unified dispatch, the output of renewable energy, and the output of conventional generating units should satisfy the supply-demand balance. Net load is defined as the difference between the load and the output of renewable energy, representing the output required from conventional generating units. The nodal marginal price should monotonically increase with net load; that is, the higher the net load, the higher the price. These constraints are converted into penalty terms in the loss function. For the supply-demand balance constraint, a penalty term in the form of squared error is constructed; for the price monotonicity constraint, a term that penalizes only violations of monotonicity is constructed. The strength of the constraints is controlled by setting a penalty coefficient. Three prediction branch networks are constructed, each containing a multi-layer fully connected neural network. The input to each branch is a concatenation of the corresponding task-specific weighted feature vector and a task-shared general feature vector. The loss function of each branch includes a prediction error term and a cross-constraint penalty term, achieving multi-task collaborative learning through joint optimization.

[0114] The load probability distribution parameters are determined through the load forecasting branch; the renewable energy output probability distribution parameters are determined by using the load probability distribution parameters as additional inputs through the renewable energy output forecasting branch; and the electricity price probability distribution parameters are determined by using both the load probability distribution parameters and the renewable energy output probability distribution parameters as additional inputs through the electricity price forecasting branch.

[0115] The load forecasting branch adopts a probabilistic forecasting architecture. The output layer outputs the probability distribution parameters of the load, including the mean and standard deviation; the standard deviation is guaranteed to be positive through an activation function. Weather forecast data and historical load patterns are used as conditional variables; this information is encoded in a task-specific weighted feature vector. Assuming the load follows a normal distribution, the uncertainty of the forecast can be quantified by outputting distribution parameters rather than point prediction values. The weighted sum of the negative log-likelihood loss and the cross-constraint penalty term is minimized. The negative log-likelihood loss measures the degree of fit between the predicted distribution and the true value, while the cross-constraint penalty term guides the model to learn predictions that conform to physical laws. For each moment in the next 24 hours, the load probability distribution parameters, including the mean and standard deviation, are output.

[0116] By treating the load probability distribution parameters as an additional feature and concatenating them with the task-specific weighted feature vector and the task-shared general feature vector for renewable energy output forecasting, this design enables renewable energy output forecasting to perceive the results and uncertainties of load forecasting, reflecting the impact of load on renewable energy absorption capacity. The renewable energy output forecasting branch also adopts a probabilistic forecasting architecture, outputting the probability distribution parameters of renewable energy output. Using meteorological forecast data, unit operating status, and load forecasting information as conditional variables, it assumes that renewable energy output follows a truncated normal distribution, with the truncated range between zero and installed capacity, to satisfy physical constraints. By minimizing the weighted sum of the negative log-likelihood loss and cross-constraint penalty term, the probability distribution parameters of renewable energy output are output for each moment in the next 24 hours.

[0117] By using load and renewable energy output probability distribution parameters as additional features, and concatenating them with the task-specific weighted feature vector and the task-shared general feature vector, electricity price forecasting can simultaneously perceive the forecast results and uncertainties of load and renewable energy output, reflecting the comprehensive impact of supply and demand on electricity prices. The electricity price forecasting branch adopts a probabilistic forecasting architecture, outputting the probability distribution parameters of electricity prices. Load forecasting information, renewable energy output forecasting information, fuel prices, and historical electricity price patterns are used as condition variables. It assumes that electricity prices follow a log-normal distribution to better capture the positive skewness and non-negative constraints of electricity prices. By minimizing the weighted sum of the negative log-likelihood loss and cross-constraint penalty terms, the probability distribution parameters of electricity prices are output for each moment in the next 24 hours.

[0118] A preset number of future scenarios are generated by sampling from three probability distribution parameters. The net load value is calculated for each scenario. The net load value and the sampled electricity price value are checked to see if they satisfy a monotonically increasing relationship. The sampled value of renewable energy output is also checked to see if it is within the range of physical constraints. Inconsistency indicators are statistically analyzed.

[0119] For each moment in the next 24 hours, samples are taken from the probability distributions of load, renewable energy output, and electricity price to generate a large number of possible future scenarios. In this embodiment, 1000 scenarios are sampled. The sampling process follows the respective probability distribution assumptions; for example, load is sampled from a normal distribution, renewable energy output from a truncated normal distribution, and electricity price from a log-normal distribution. For each scenario, the net load value at each moment is calculated, which is the load sample value minus the renewable energy output sample value. The net load reflects the output demand that needs to be provided by conventional units. The correlation between net load and electricity price is checked to see if it satisfies a monotonically increasing relationship. Specifically, the rank correlation coefficient between the net load sequence and the electricity price sequence in each scenario is calculated. If the correlation coefficient is lower than a set threshold (e.g., 0.5), the scenario is determined to violate the monotonicity constraint and does not conform to the price formation law of the electricity market. The renewable energy output sample value is checked to see if it is within the physical constraint range, i.e., between zero and the installed capacity; if it exceeds this range, the scenario is determined to violate the physical constraint. The number of scenarios that violate monotonicity or physical constraints is counted, and the inconsistency index is calculated as the proportion of scenarios that violate constraints to the total number of scenarios.

[0120] When the inconsistency index exceeds the third threshold, adjust the probability distribution parameters or truncate the distribution tails that exceed the physical constraints, and repeat the physical consistency check until the inconsistency index drops below the third threshold or reaches the preset number of iterations, and determine the corrected probability distribution parameters.

[0121] A third threshold is set; in this embodiment, it is set to 0.05. If the inconsistency index exceeds this threshold, it indicates insufficient physical consistency of the prediction results, requiring correction. Different adjustment strategies are adopted based on the type of constraint violation. For scenarios violating monotonicity constraints, the relationship between net load and electricity price is analyzed, and the uncertainty range of the electricity price distribution parameters is adjusted to better conform to the monotonically increasing law. For scenarios violating physical constraints, the distribution tails exceeding the physical constraints are truncated, or the distribution parameters are adjusted to make their centers further away from the constraint boundaries. The sampling, verification, and parameter adjustment process is repeated until the inconsistency index drops below the third threshold. To avoid over-iteration, a maximum number of iterations is set; in this embodiment, it is set to 10. After iterative correction, the final probability distribution parameters are determined. These parameters retain the uncertainty information of the prediction while satisfying the physical consistency requirements.

[0122] The corrected probability distribution parameters are aligned along the time axis, and the confidence intervals and covariance matrices at each time point are calculated. A structured prediction dataset containing the predicted mean, confidence intervals, covariance matrix, and inconsistency index is constructed as the probability distribution prediction result for future periods.

[0123] Finally, the corrected probability distribution parameters are transformed into a structured prediction dataset, including the prediction mean, confidence intervals, covariance matrix, and inconsistency index. The prediction mean provides the most probable predicted value, the confidence interval quantifies the range of uncertainty in the prediction, the covariance matrix reflects the correlation and uncertainty propagation among multiple predictors, and the inconsistency index reflects the physical reliability of the prediction results. Compared to traditional point predictions, this structured output provides decision-makers with more complete information, supporting risk assessment and robust decision-making. For example, when formulating scheduling plans, decision-makers can set reserve capacity based on confidence intervals, assess the joint risk of multiple factors deviating from the predicted values ​​simultaneously based on the covariance matrix, and judge the reliability of the prediction results based on the inconsistency index.

[0124] The working principle and beneficial effects of the above technical solution are as follows:

[0125] This invention accurately captures the dynamic coupling patterns of multi-source heterogeneous data by explicitly modeling the spatiotemporal dependencies between data sources and performing time alignment compensation. It models the causal dependency chain of load, renewable energy output, and electricity price through a cascaded probabilistic prediction mechanism, conveying prediction information and uncertainty while avoiding logical contradictions in independent predictions. By introducing physical constraints in both the training and prediction phases through differentiable penalty terms and posterior verification mechanisms, it significantly reduces prediction results that violate physical laws. It provides complete uncertainty quantification information by outputting probability distribution parameters, confidence intervals, and covariance matrices. Finally, it balances knowledge transfer between tasks and specific performance optimization through a task-oriented attention mechanism. This invention achieves high-precision collaborative prediction of multiple factors in the power market, providing reliable support for power system dispatching, trading decisions, and risk assessment.

[0126] A multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, outputting a set of day-to-day and real-time trading strategies that are robust under different risk preferences, including:

[0127] Based on the constraint boundaries of the decision variables in the stochastic optimization model, random sampling is performed in the multidimensional policy space to initialize the initial position and initial velocity of each particle in the particle swarm optimization algorithm, obtaining the initial particle state vector set; the initialization formulas for the initial position and initial velocity are as follows:

[0128] ;

[0129] ;

[0130] in, Indicates the first The initial position vector of each particle represents a candidate trading strategy; and These represent the lower and upper bound vectors of the decision variables in a stochastic optimization model, respectively. This represents a vector of random numbers whose values ​​are uniformly distributed between [0,1]. Indicates the first The initial velocity vector of each particle;

[0131] The position vector of each particle in the initial particle state vector set is substituted into the stochastic optimization model for simulation evaluation. The objective function value of maximizing expected return and the objective function value of risk are calculated for each state position, and a multi-objective fitness evaluation matrix containing the two-dimensional objective values ​​associated with each particle is determined. The formula for calculating the multi-objective fitness evaluation matrix is ​​as follows:

[0132] ;

[0133] in, Indicates the first During the nth iteration The multi-objective fitness vector of each particle; Indicates the first The particles are in the iteration number of Position vector at time; This represents the expected return objective function in a stochastic optimization model; This represents the objective function for minimizing the value at risk in a stochastic optimization model; Represents the transpose of a matrix;

[0134] Calculate the dominance relationships among the particles in the multi-objective fitness evaluation matrix, update the matrix based on the dominance relationships, and determine the individual historical best position and global best position for each particle; the update formula for the individual historical best position is:

[0135] ;

[0136] in, Indicates the first Individual particles The individual's historical optimal position vector after the next iteration; This represents the Pareto dominance operator, i.e. The current state multi-objective fitness vector strictly dominates the previous generation state multi-objective fitness vector; Indicates the first The vector of the individual historical best position of each particle in the previous generation;

[0137] Based on the individual historical best position and the global best position, the positional migration state of each particle after a set number of iterations in the multidimensional policy space is continuously calculated to determine the Pareto optimal solution set; the velocity and position iterative update formulas are as follows:

[0138] ;

[0139] ;

[0140] in, , They represent the first The velocity vector and position vector of each particle in the next iteration; Indicates the first The particles are in the iteration number of The velocity vector at time; Indicates the inertia weighting coefficient; For individual cognitive learning factors; For social group learning factors; It represents the empirical random distribution coefficient of an individual generated independently in the interval [0,1], and is used to characterize the degree to which the particle randomly references its historical best position when searching in the multidimensional policy space; It represents the empirical random distribution coefficient of the population generated independently in the interval [0,1], which is used to characterize the degree to which the particle randomly listens to the current global optimal position of the system when searching in the multidimensional policy space; The globally optimal position vector is provided by the multi-objective optimization-guided parameter set;

[0141] Based on the Pareto optimal solution set and the pre-introduced multi-level risk preference coefficients of users, the target position vector is determined by a linear weighted matching method. This target position vector is then mapped to a set of day-to-day and real-time trading strategies that are robust under different risk preferences. Specifically, the linear weighted matching method for determining the target position vector is implemented using the following risk strategy matching formula:

[0142] ;

[0143] in, Indicates the first The optimal strategy position vector (target position vector) extracted from the Pareto optimal solution set under a certain user risk preference coefficient. This represents the optimization operator, which searches within the Pareto optimal solution set for a specific position vector that maximizes the value of the weighted objective function within the parentheses. ; This represents the set of Pareto optimal solutions received. The user introduced Risk preference coefficient, and [0,1]; Represents the candidate position vector The expected return objective function value after mean-variance normalization; Represents the candidate position vector The value at risk objective function value after mean-variance normalization.

[0144] The working principle and beneficial effects of the above technical solution are as follows:

[0145] This invention first initializes the particle swarm by randomly sampling within a multidimensional policy space based on the constraint boundaries of the decision variables in the stochastic optimization model. Then, the position vector of each particle is substituted into the stochastic optimization model for simulation evaluation, and the expected return objective function value and risk value objective function value are calculated respectively, constructing a multi-objective fitness evaluation matrix containing the two-dimensional objective values ​​of all particles. Next, the individual historical best position and global best position of each particle are updated by comparing Pareto dominance relationships. The dominance relationship is determined by comparing two multi-objective fitness vectors; if the current state is not inferior to the historical state in all objective dimensions and is strictly superior to the historical state in at least one objective, then the current state is considered to dominate the historical state. Finally, iterative calculations are performed according to the velocity update formula and the position update formula, where the inertia weight coefficient... It is typically set to a decreasing value between 0.4 and 0.9 to balance global exploration and local development capabilities, and the individual cognitive learning factor. Social group learning factors Typically, the value is set to 1.5-2.0. After a set number of iterations, the algorithm converges and outputs a Pareto optimal solution set. There is no dominance relationship between any two solutions in this set, which represents the optimal trade-off frontier between expected return and risk control. Finally, based on the user-preset multi-level risk preference coefficients, the target strategy is extracted from the Pareto optimal solution set using a linear weighted matching method.

[0146] For example, suppose a virtual power plant needs to formulate a 24-hour electricity trading strategy for the following day, including day-ahead market bids and real-time market adjustments. This virtual power plant aggregates 100MW of wind power, 50MW of photovoltaic power, and 30MWh of energy storage systems, and needs to optimize day-ahead and real-time market trading decisions while meeting equipment operating constraints.

[0147] The particle swarm size is set to 100 particles, and the maximum number of iterations is 200. Decision variables include daily reported electricity consumption and real-time adjusted electricity consumption for 24 time periods, resulting in a 48-dimensional decision space. A lower limit for the decision variables is set based on equipment capacity constraints. (Unit: MW), Lower Limit .

[0148] The particle positions are initialized using uniform random sampling. For example, the initial position of the first particle is: The initial velocity of all particles is set to zero vector. .

[0149] Set algorithm parameters: Inertia weight The individual cognitive learning factor decreased linearly from 0.9 to 0.4. Social group learning factors .

[0150] The initial positions of 100 particles were substituted into a stochastic optimization model, and Monte Carlo simulations were performed based on 500 wind and solar power output scenarios and electricity price scenarios. The multi-objective fitness of the first particle was calculated: The expected return is 12,500 yuan, and the conditional value at risk (CVaR) is 3,200 yuan.

[0151] In the first iteration, since there is no historical record, the individual best position of each particle is its initial position: By comparing the fitness vectors of all particles, the non-dominated solutions that constitute the global optimal set are identified. For example, if particle 1 has a reward of 12,500 yuan and a risk of 3,200 yuan, and particle 2 has a reward of 13,000 yuan and a risk of 3,500 yuan, then they do not dominate each other; if particle 3 has a reward of 12,000 yuan and a risk of 3,300 yuan, then it is dominated by particle 1.

[0152] In the In the next iteration, random coefficients are generated. Randomly select a position from the globally optimal set as (For example, the non-dominated solution with the highest yield). Calculate the velocity update of the first particle: Location update: If the updated position exceeds the constraint boundary, then it will be restricted to... Within the range.

[0153] Repeat the above steps for 200 iterations. As the iterations proceed, particles gradually converge toward the Pareto front. After the 200th iteration, the global optimal set contains 25 non-dominated solutions, forming the Pareto optimal solution set. The expected returns of these solutions range from [11000, 15000], and the corresponding risk values ​​range from [4500, 2000], exhibiting a negative trade-off between returns and risks.

[0154] Users can set three risk preference coefficients: (Conservative type) (Balanced type) (Aggressive approach). After normalizing the mean and variance of each solution in the Pareto optimal solution set, the weighted objective function value is calculated.

[0155] For a conservative strategy, the conclusion is... The corresponding expected return is 11,800 yuan and the risk value is 2,200 yuan. The current reporting strategy is to maintain a low trading volume in all time periods, and make minor adjustments in the real-time market.

[0156] For a balanced strategy, the conclusion is... The corresponding expected return is 13,200 yuan and the risk value is 3,100 yuan. The current reporting strategy is to moderately report electricity sales during periods of higher electricity prices, and make moderate adjustments in real time based on the deviation between wind and solar power output, so as to achieve a balance between returns and risks.

[0157] For aggressive strategies, the conclusion is... The corresponding expected return is 14,500 yuan and the risk value is 4,200 yuan. The current reporting strategy is to submit a large number of electricity sales orders during periods of high electricity prices, and the real-time market will make significant adjustments based on the actual output.

[0158] The final output consists of three sets of trading strategies, each containing the day-ahead reported electricity volume and real-time adjusted electricity volume for 24 time periods, which virtual power plant operators can choose to execute based on the market environment and risk appetite of the day.

[0159] This invention employs a multi-objective particle swarm optimization algorithm to solve stochastic optimization models, simultaneously addressing the conflicting optimization objectives of maximizing expected return and minimizing risk value. This avoids the subjectivity issues of traditional single-objective optimization methods that require manually pre-setting weight coefficients. The particle swarm algorithm features strong global search capabilities, fast convergence speed, and simple parameter settings. Compared to evolutionary algorithms such as genetic algorithms, it offers higher computational efficiency when handling continuous optimization problems. Furthermore, through dynamic adjustment of inertia weights and the introduction of random learning factors, it effectively avoids getting trapped in local optima. The output Pareto optimal solution set provides decision-makers with a series of non-dominated trading strategy choices. Each strategy represents an optimal trade-off between returns and risks, allowing decision-makers to flexibly choose based on the actual market environment and their own risk tolerance. By introducing multi-level risk preference coefficients and combining them with a linear weighted matching method, the system can automatically select the strategy that best matches the user's risk preferences from the Pareto frontier, significantly improving the adaptability and robustness of the trading strategy. Furthermore, the day-ahead and real-time trading strategy sets generated by this method can cope with various stochastic factors such as electricity market price fluctuations and uncertainties in renewable energy output, providing a scientific decision support tool for virtual power plants to participate in electricity market transactions, and helping to improve the level of renewable energy consumption and the economic efficiency of power system operation.

[0160] The transaction execution module includes:

[0161] The strategy compilation module is configured to compile the received preferred strategy parameters into standardized trading instruction messages that conform to the communication protocol of the power trading platform. Specifically:

[0162] Receive a set of preferred strategies output from the intelligent prediction module, where each strategy exists in the form of a structured set of parameters, including but not limited to trading instruments, trading hours, order volume curves, order price curves, effective time windows, trigger and cancellation conditions, etc.

[0163] The strategy compilation module has pre-built communication protocol templates and message format definitions that are fully compatible with the official API specifications of the target power trading platform;

[0164] During compilation, the strategy compilation module performs compliance checks on the input strategy parameters to ensure they comply with trading rules. Then, based on the strategy type (such as limit orders or conditional orders), it selects the corresponding protocol template and maps, fills, and formats the strategy parameters according to the field order, data type (such as integer, floating-point, or string) and encoding rules specified in the template. This generates one or more standardized text or binary messages containing complete trading instruction information. These messages can be directly parsed and executed by the trading platform API. At the same time, a unique session identifier and timestamp are attached to each generated message for subsequent tracking and tracing.

[0165] The contract execution module is configured to have dynamically configurable smart contract scripts deployed internally. The smart contract scripts encapsulate complete transaction logic for order placement, cancellation, and confirmation, and can monitor and judge multiple conditions such as preset price, time, trading volume, and price difference of related products in parallel based on real-time access to market data streams.

[0166] The condition triggering module is configured to automatically invoke standardized trading instruction messages when market data is detected to meet the triggering conditions of any smart contract, execute the corresponding trading operation through the power trading platform API, and simultaneously generate an immutable operation log with timestamps and execution environment hash values.

[0167] The technical effects of the above solution are as follows: the strategy compilation module can automatically transform complex optimization strategies into standardized executable instructions, thereby ensuring the accurate transmission of trading intentions and platform compatibility; the contract execution module uses dynamically configurable smart contract scripts to encapsulate complex trading logic and achieve parallel and real-time monitoring of multiple market conditions, thus providing a flexible and reliable decision-making core for automated execution; and the condition triggering module realizes the precise linkage and automatic execution of market conditions and trading instructions, and ensures the immediacy, auditability, and immutability of each trading operation by generating operation logs with verifiable hash values, thereby improving execution efficiency and accuracy while comprehensively strengthening risk control and compliance transparency in the trading process.

[0168] The knowledge management module includes:

[0169] The post-trade analysis module is configured to automatically perform the following analyses at three levels: daily, time-segment, and single transaction, based on transaction execution logs and actual market clearing data:

[0170] Attribution analysis is performed on the deviation between the predicted values ​​and the actual values ​​output by the coupled prediction model;

[0171] Compare the expected returns of the trading strategy with the actual risk-adjusted returns and attribute the performance.

[0172] The knowledge graph update module is configured to structure and map the attribution conclusions, key performance indicators and corresponding market scenario characteristics obtained from the retrospective analysis module into entity nodes, attribute relationships and confidence labels in the knowledge graph.

[0173] Specifically, when the confidence level of the relationship between a specific scenario and a strategy pattern reaches a preset threshold, a model optimization instruction is triggered.

[0174] Update the driver module to respond to model optimization commands, convert high-confidence knowledge graph triple data into an incremental learning sample set for coupled prediction models and constraint rules for policy optimization, and drive targeted iterative updates of coupled prediction models and policy rules.

[0175] The technical effects of the above solution are as follows: the review analysis module performs multi-level and automated quantitative attribution and performance evaluation of the trading process, providing accurate diagnostic basis for strategy optimization; the graph update module transforms the review conclusions and market scenario characteristics into a structured and reasonable knowledge graph, thereby realizing the systematic accumulation of experience and rules; and the update-driven module drives the continuous and targeted iterative optimization of the prediction model and trading strategy by transforming high-confidence knowledge into model training samples and strategy constraints, thus forming a complete closed loop from practical evaluation to knowledge accumulation and system evolution, significantly improving the self-learning ability and long-term adaptability of the entire decision-making system.

[0176] Specifically, this embodiment also proposes a power trading auxiliary decision-making method based on multi-source data and intelligent algorithms, which is applied to a power trading auxiliary decision-making system based on multi-source data and intelligent algorithms. The method includes the following steps:

[0177] Real-time acquisition and fusion of multi-source data: Market, meteorological, and unit operation data are collected through a cloud-edge collaborative architecture, and after cleaning, verification, time alignment and hash identification, they are stored in a distributed storage layer;

[0178] Coupled Cooperative Prediction and Strategy Generation: A pre-trained coupled prediction model is invoked to synchronously predict future market factors. Combined with market rules and risk constraints derived from a knowledge graph, a multi-objective optimization algorithm is used to generate multiple candidate trading strategies, which are then simulated and compared to output a recommended strategy set.

[0179] Smart contract-driven automated trade execution: The recommended strategy is transformed into a smart contract script with trigger conditions, which is deployed in the trade execution module. When market conditions are met, the order placement operation is automatically executed, and the execution status and transaction results are monitored throughout the process.

[0180] Closed-loop review and knowledge iteration of the entire trading process: Based on actual transaction data and market results, automated review and analysis are conducted on prediction accuracy, strategy effectiveness, and execution compliance. The analysis results are structured and fed back to the knowledge graph and coupled prediction model to achieve self-learning and continuous optimization of the system.

[0181] The technical effects of the above-mentioned solution are as follows: the cloud-edge collaborative architecture enables real-time reliable fusion and efficient processing of multi-source data, thereby providing a high-quality data foundation for decision-making; by leveraging the strategy generation of coupled collaborative prediction and fusion market rules, it achieves accurate quantification of complex market uncertainties and multi-objective robust strategy optimization; through automated execution driven by smart contracts, it ensures the accurate, timely, and compliant implementation of trading strategies, thereby reducing the risks and delays of manual operation; and through the closed-loop review and knowledge iteration mechanism, it can continuously transform practical experience into system knowledge, thereby driving the autonomous evolution of prediction models and decision-making strategies, thus comprehensively improving the intelligence level, adaptability, and long-term profitability stability of power trading decisions.

[0182] The pre-trained coupled prediction model is invoked to simultaneously predict future market factors, specifically including:

[0183] Based on a multi-task learning framework, a shared feature layer is used to extract spatiotemporal correlation features from multi-source data.

[0184] The probability distributions of load, renewable energy output, and electricity price are predicted by outputting them through dedicated network branches.

[0185] The forecast results and uncertainty range are integrated with current policy constraints, unit physical limitations, and contractual obligations retrieved from the knowledge graph;

[0186] By taking the maximization of expected returns and the control of risk value as optimization objectives, a constrained swarm intelligence algorithm is used to search the policy space and output the optimal policy frontier for decision-makers to choose from.

[0187] The technical effects of the above-mentioned solution are as follows: By sharing feature extraction through a multi-task learning framework, it can capture and utilize the deep spatiotemporal correlations and dependencies among key elements in the power market, thereby improving prediction accuracy and consistency. By outputting probability distributions and uncertainty intervals, it achieves a quantitative representation of market risk. Furthermore, by integrating the prediction results with real-time rules, physical and contractual constraints extracted from the knowledge graph, it ensures the compliance and feasibility of the strategy. Through multi-objective optimization, it finds an equilibrium solution between returns and risks, providing decision-makers with a series of scientific, comprehensive, and operable robust trading strategies. This enables intelligent, refined, and risk-controllable auxiliary decision-making in complex and uncertain market environments.

[0188] The entire process of execution and transaction feedback is monitored, including:

[0189] Parameterize the price curve, volume, and trigger logic (price, time, and related commodity signals) in the strategy;

[0190] Generate smart contract code and deploy it in a trusted execution environment, with the contract state connected to real-time market data streams;

[0191] When market data is detected to meet the contract trigger conditions, the contract automatically calls the trading platform API to send compliant order placement instructions;

[0192] Record the entire lifecycle log of the contract from creation and triggering to completion, and calculate the hash value for evidence storage.

[0193] The technical effects of the above solution are as follows: By using parameterized smart contracts, complex trading strategies are transformed into reliable code that can be executed automatically. Through real-time connection with market data streams, precise monitoring and instantaneous response to trigger conditions are achieved, thereby ensuring the timeliness and accuracy of transaction execution. At the same time, by recording full lifecycle logs and supplementing them with hash-based evidence storage, a complete, transparent, and tamper-proof transaction audit and traceability chain can be established. This not only improves the efficiency of automated execution and reduces operational risks, but also greatly enhances the compliance, credibility, and post-transaction verifiability of the transaction process.

[0194] Based on actual transaction data and market results, an automated post-mortem analysis is conducted to assess forecast accuracy, strategy effectiveness, and execution compliance. This includes:

[0195] Deviation analysis is performed between predicted and actual values ​​to identify sources of systematic error and adjust model features and structure.

[0196] Compare the simulated returns of the strategy with the risk-adjusted returns after actual implementation to analyze whether the performance difference stems from forecast bias or unexpected market events.

[0197] The new market patterns, effective strategy models, and lessons learned during this trading cycle will be constructed into new entities, relationships, and rules in a knowledge graph, and their confidence levels will be labeled.

[0198] When the confidence level of new knowledge exceeds a set threshold, the incremental learning process of the coupled prediction model and the rule base update of the policy optimizer are triggered.

[0199] The technical effects of the above-mentioned technical solution are as follows: Through multi-layered automated review analysis, a comprehensive diagnosis can be achieved from prediction deviation attribution to strategy performance evaluation, thereby providing a precise basis for system optimization. By transforming the new rules, patterns, and lessons learned from the analysis into a structured knowledge graph, the systematic accumulation and reasonable storage of experience are realized. Furthermore, by triggering incremental learning and rule updates through confidence thresholds, the coupled prediction model and strategy optimization can dynamically absorb high-value experience, ultimately driving the entire decision-making system to form a closed loop of continuous self-improvement, thereby enhancing the adaptability to market changes, the long-term effectiveness of strategies, and the overall level of decision-making intelligence.

[0200] The preferred embodiments of the present invention will be described in detail below, but the scope of protection of the present invention is not limited thereto:

[0201] Example 1: New energy power plants participate in the joint bidding of the day-ahead market and the real-time market

[0202] A wind farm participates in the electricity spot market and makes day-ahead and real-time trading decisions through this system.

[0203] The specific process is as follows:

[0204] Data fusion:

[0205] The data fusion module uses a cloud-edge collaborative architecture to collect real-time weather forecasts (wind speed, wind direction), turbine operating status (turbine availability, active power), day-ahead market historical clearing prices, load forecast data, and grid congestion information for the wind farm. After quality verification and alignment, the data is stored in a time-series database.

[0206] Intelligent prediction and strategy generation:

[0207] Based on the above data, the coupled prediction model in the intelligent prediction module simultaneously outputs the probability distribution of the unified dispatch load in the wind farm's location area, the probability prediction of wind power output, and the probability range of the nodal marginal price (LMP) for the next 24 hours.

[0208] The strategy optimization module constructs a multi-objective optimization model based on the user's "medium risk appetite" and "monthly revenue target constraint," combined with constraints such as "real-time market penalty rules" and "unit ramp-up restrictions" retrieved from the knowledge graph. Through a multi-objective particle swarm optimization algorithm, it obtains a set of day-ahead market declaration curves (containing declared electricity volumes at different prices) and multiple real-time market adjustment strategy plans (e.g., increasing generation when the real-time electricity price is above a certain threshold and decreasing generation when it is below a certain threshold).

[0209] Transaction execution:

[0210] The trade execution module compiles the selected day-ahead order curves into standard messages and automatically submits them to the trading platform at the specified time in the day-ahead market. Simultaneously, it encapsulates real-time market adjustment strategy plans into multiple smart contracts. During real-time operation, the system continuously monitors real-time LMP and turbine status; when conditions are met, it automatically triggers the corresponding smart contracts, submitting adjusted quotes to the real-time market via API.

[0211] Knowledge Management and Iteration:

[0212] After the day's trading concluded, the knowledge management module automatically conducted a post-mortem analysis. It discovered that during the afternoon peak load period, actual wind power output was lower than predicted. However, due to soaring real-time electricity prices, the system-triggered high-price power reduction strategy actually avoided performance penalties for insufficient power, resulting in overall higher-than-expected returns. The post-mortem analysis stored the pattern of a positive correlation between wind power forecasting error and real-time electricity prices under specific weather patterns (such as afternoon turbulence), along with effective strategies for dealing with this scenario, as high-confidence knowledge in the knowledge graph. Subsequently, when the coupled forecasting model encounters similar weather characteristics, it will strengthen the estimation of output uncertainty, and the strategy optimization module will be more inclined to generate strategies that combine conservative day-ahead reporting with flexible real-time responses.

[0213] Example 2: Risk-reward balance of conventional thermal power units under complex market rules

[0214] A coal-fired power plant faces stringent emission constraints, complex bilateral contract breakdowns, and ancillary services market opportunities, and needs to utilize this system for refined transactions.

[0215] The specific process is as follows:

[0216] Data fusion:

[0217] The system integrates data from multiple sources, including enterprise internal management systems (fuel inventory, unit maintenance plans, signed bilateral contracts), spot market operators (clearing prices, ancillary service demands), and environmental protection departments (emission limits), to form a unified decision-making view.

[0218] Intelligent prediction and strategy generation:

[0219] The coupled prediction model outputs probabilistic predictions of future load, electricity price, and frequency regulation ancillary service demand. The knowledge graph reasoning engine of the strategy optimization module dynamically parses complex rules such as "time-of-use emission intensity constraints," "rigid requirements of bilateral contract decomposition," and "restrictions on unit operating status for participating in the frequency regulation market."

[0220] The user sets the goal to "maximize profits while controlling market risk exposure, provided that contract performance is guaranteed." The strategy optimization module constructs a stochastic optimization model that simultaneously considers the electricity market, frequency regulation market, and carbon emission costs. Using a multi-objective particle swarm optimization algorithm, it solves for a set of Pareto optimal strategies, including: electricity allocation schemes between the energy market and frequency regulation market at different times, and suggestions for adjusting the bidding curve considering carbon emission costs.

[0221] Transaction execution:

[0222] The system decomposes the overall optimal strategy into electricity market bidding instructions and ancillary services market application instructions. The transaction execution module submits applications in an orderly and automatic manner according to the time windows of different markets. The smart contract engine simultaneously monitors energy market prices and frequency regulation clearing signals, and automatically executes cross-market strategy switching when conditions permit (for example, when the frequency regulation compensation price far exceeds the energy market price, it automatically adjusts the unit mode to strive for winning the frequency regulation bid).

[0223] Knowledge Management and Iteration:

[0224] The post-mortem analysis revealed that, at a certain point in time, the strategy, overly focused on high returns in the frequency regulation market, led to frequent unit adjustments. While frequency regulation compensation was achieved, this increased unit wear and fuel costs, resulting in suboptimal overall efficiency. The knowledge management module incorporated this lesson—that "the overall cost under the frequent adjustment model was underestimated"—along with the market characteristics of the time (such as price volatility and the duration of frequency regulation demand), into the knowledge graph as new rules and relationships. Subsequently, when the strategy optimization module builds optimization models in similar scenarios, it automatically introduces a "virtual cost" item related to the adjustment frequency, thereby generating a strategy that better balances economic efficiency and equipment lifespan.

[0225] Example 3: Risk Hedging Strategy for Electricity Sales Companies to Act as Agents for Users in Electricity Purchase

[0226] An electricity sales company that acts as an agent for multiple industrial and commercial users to participate in the spot market for electricity purchases needs to manage the risk of electricity price fluctuations and ensure the stability of users' electricity costs.

[0227] The specific process is as follows:

[0228] Data fusion:

[0229] The system integrates users' load curves (historical and forecast), spot market electricity price data, forward contract price information, and macroeconomic indicators.

[0230] Intelligent prediction and strategy generation:

[0231] The coupled prediction model focuses on high-precision probability distribution prediction of nodal marginal prices (LMP), including peak, high, flat, and low periods. The knowledge graph stores the electricity consumption characteristics, risk tolerance (e.g., some users are very sensitive to extreme price fluctuations), and company risk management policies of different user types.

[0232] The strategy optimization module aims to "minimize the overall electricity purchase cost for users and control the risk of cost fluctuations" by searching the decision space for the optimal combination of spot market electricity purchase ratio and hedging positions with financial derivatives such as forward / futures.

[0233] Transaction execution:

[0234] The transaction execution module coordinates operations between the financial derivatives trading platform and the electricity spot trading platform based on the selected strategy. For example, when the system predicts that extremely high prices may occur in a future week, the smart contract will automatically execute a buy call spread contract in the financial market to hedge the procurement risk in the spot market. In the spot market, it automatically generates time-of-use electricity purchase orders based on time-of-use forecasted electricity prices and user load curves.

[0235] Knowledge Management and Iteration:

[0236] A quarter later, the knowledge management module conducts a periodic in-depth review. Its analysis reveals that during seasons of high renewable energy generation and low system marginal costs, the "high-proportion spot trading + dynamic hedging" strategy yields significant excess returns. However, during periods of sharp fuel price fluctuations, the risk-adjusted return of this strategy is actually lower than that of a "long-term contract-based" strategy. The system extracts these mappings between market conditions (such as energy structure and fuel cost volatility) and optimal strategy types, forming high-confidence scenario-strategy pattern pairs, and updates them to the knowledge graph. In the future, when the market condition identification module determines that the current environment is highly similar to a historical scenario, the system will prioritize recommending the strategy pattern with the highest correlation to that scenario and make personalized fine-tuning, thereby continuously improving the adaptability and robustness of decision-making.

[0237] The above three embodiments illustrate, from the perspectives of three typical market players—new energy power plants, traditional thermal power plants, and electricity sales companies—how this invention achieves intelligent, automated, and continuously optimized decision-making in electricity trading through a complete process of data fusion, intelligent prediction, strategy optimization, automatic execution, and closed-loop learning, effectively addressing the core challenges in different scenarios.

[0238] Working Principle: Through a cloud-edge collaborative data fusion module, multi-source heterogeneous data from the market, power grid, meteorology, and generating units are aggregated and standardized in real time, providing a unified and high-quality time-series data foundation for decision-making. The core intelligent forecasting module employs a coupled forecasting model based on multi-task learning, simultaneously outputting the probability distribution of future load, renewable energy output, and nodal electricity prices, accurately quantifying market uncertainty. Simultaneously, combining market rules and user risk preferences with knowledge graph reasoning, the strategy optimizer uses a multi-objective optimization algorithm to search in multiple simulated market scenarios, generating a set of robust trading strategy candidates that balance returns and risks. The trading execution module compiles the selected strategy into dynamically triggerable smart contracts, directly connecting to the trading platform via API to automate order placement and cancellation operations when conditions are met, ensuring accurate and timely strategy execution. The knowledge management module automates the review and performance attribution of each transaction, structurally storing experiences and lessons learned in a knowledge graph. This drives targeted iterative optimization of the predictive model and strategy rules, enabling the system to continuously evolve. Through collaborative prediction and strategy optimization, this invention helps decision-makers capture opportunities and manage risks in complex and ever-changing markets. By automating the entire process, it greatly improves the efficiency and accuracy of trading operations, thereby reducing human error and response delays. Furthermore, through a closed-loop self-learning mechanism, the system can continuously accumulate experience and adapt to changes in market rules and the environment, thus continuously enhancing its decision-making capabilities.

[0239] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0240] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A power trading auxiliary decision-making system based on multi-source data and intelligent algorithms, characterized in that, include: The data fusion module is configured to collect multi-source data from electricity spot market operators, meteorological departments, and enterprise internal management systems, and to clean, align, and timestamp the collected multi-source data, generating unique hash values ​​for data entries and storing them in a distributed time-series database. The multi-source data includes at least market clearing data, grid boundary conditions, weather forecast data, generator unit operating parameters, and historical transaction data. The intelligent forecasting module is configured to access multi-source data processed by the data fusion module. Based on a coupled forecasting model built using deep learning and reinforcement learning, it is used to simultaneously forecast the unified dispatch load, renewable energy output, and nodal marginal electricity price for a specific future period. It also integrates a market rule reasoning engine based on a knowledge graph and a strategy optimizer based on a multi-objective particle swarm optimization algorithm. Based on the forecast results, user-defined risk preferences, and cost constraints, it generates and compares multiple alternative day-ahead and real-time trading quotation strategies. The transaction execution module is configured to interface with the power trading platform API, receive the preferred strategies output by the intelligent prediction module, convert the strategy instructions into executable standard messages, and automatically execute order placement, order cancellation and transaction confirmation operations when preset market conditions are met through the embedded smart contract engine, and track market transaction status and position status in real time. The knowledge management module is configured to perform automatic review and analysis on a daily, time-period, and transaction basis. By comparing the actual clearing results with the prediction strategy, it quantitatively evaluates the performance of the strategy and stores the review conclusions in the knowledge graph in association with key market scenarios for iterative optimization of the prediction model and strategy rules.

2. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 1, characterized in that, The data fusion module includes: The data acquisition module is deployed on the power plant's edge computing node and is configured to collect and preprocess unit operation data and station meteorological data in real time, and upload them to the cloud data center through an encrypted channel. The quality verification module is configured to use a data anomaly detection model based on statistical rules and machine learning to verify the integrity, rationality and consistency of collected multi-source data, and to repair missing and abnormal multi-source data using a spatiotemporal correlation interpolation algorithm.

3. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 1, characterized in that, The intelligent prediction module includes: The coupled prediction module is configured to build the coupled prediction model based on a multi-task learning framework. It uses a shared feature extraction layer to encode the spatiotemporal correlation features of multi-source data and outputs the probability distribution prediction results of future cycles simultaneously through load prediction branch, new energy output prediction branch and electricity price prediction branch respectively. The strategy optimization module is configured to receive probability distribution prediction results and combine them with real-time market rule constraints and user risk preferences obtained from the knowledge graph inference engine to construct an optimization problem with the goal of maximizing expected returns and minimizing risk value. The multi-objective particle swarm optimization algorithm is used to solve the problem and output a set of day-ahead and real-time candidate pricing strategies.

4. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 3, characterized in that, A shared feature extraction layer is used to encode the spatiotemporal correlation features of multi-source data, and the probability distribution prediction results for future cycles are simultaneously output through load forecasting, renewable energy output forecasting, and electricity price forecasting branches, including: Calculate the mutual information value and time lag correlation coefficient between each data source in the multi-source data. When the mutual information value exceeds the first threshold, it is determined that there is a dependency relationship. The time delay is determined according to the peak position of the time lag correlation coefficient. Construct a directed spatiotemporal dependency graph, where nodes represent data sources, the weight of the edges is the normalized mutual information value, and the direction of the edges is determined by the time lag order. Multi-source data are weighted and fused according to the edge weights in the directed spatiotemporal dependency graph. The fusion weight is proportional to the edge weight. Time alignment compensation is performed on data source pairs whose edge weights exceed the second threshold to generate fused feature representations. The fused feature representation is input into the shared feature extraction layer to generate task-specific weighted feature vectors for three tasks: load forecasting, renewable energy output forecasting, and electricity price forecasting, as well as a task-shared general feature vector. Based on the supply and demand balance of the power system and the electricity price formation mechanism, the constraints are transformed into differentiable penalty terms. According to the specific weighted feature vectors of each task and the common feature vectors shared by the tasks, the load forecasting branch, the new energy output forecasting branch, and the electricity price forecasting branch with cross-constraint penalty terms are determined. The load probability distribution parameters are determined through the load forecasting branch; the renewable energy output probability distribution parameters are determined through the renewable energy output forecasting branch, using the load probability distribution parameters as additional inputs; and the electricity price probability distribution parameters are determined through the electricity price forecasting branch, using both the load probability distribution parameters and the renewable energy output probability parameters as additional inputs. A preset number of future scenarios are generated by sampling from three probability distribution parameters. The net load value is calculated for each scenario. The net load value and the sampled electricity price value are checked to see if they satisfy a monotonically increasing relationship. The sampled value of new energy output is also checked to see if it is within the physical constraints. Inconsistency indicators are statistically analyzed. When the inconsistency index exceeds the third threshold, adjust the probability distribution parameters or truncate the distribution tails that exceed the physical constraints, and repeat the physical consistency check until the inconsistency index drops below the third threshold or reaches the preset number of iterations, and determine the corrected probability distribution parameters. The corrected probability distribution parameters are aligned along the time axis, and the confidence intervals and covariance matrices at each time point are calculated. A structured prediction dataset containing the predicted mean, confidence intervals, covariance matrix, and inconsistency index is constructed as the probability distribution prediction result for future periods.

5. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 3, characterized in that, The strategy optimization module executes the following process: The probability distribution prediction results output by the coupled prediction module are transmitted to the strategy optimization module as uncertainty quantification input. The strategy optimization module generates multiple future market scenarios that conform to the distribution based on probability distribution prediction results through simulation. In each simulation scenario, a stochastic optimization model is constructed by combining scenario-based constraint rules obtained in real time from the knowledge graph inference engine and the user's preset risk preferences, with the goal of maximizing expected utility. A multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, and a set of day-to-day and real-time trading strategies with robustness under different risk preferences is output.

6. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 5, characterized in that, A multi-objective particle swarm optimization algorithm is used to solve the stochastic optimization model, outputting a set of day-to-day and real-time trading strategies that are robust under different risk preferences, including: Based on the constraint boundary of the decision variables in the stochastic optimization model, random sampling is performed in the multidimensional policy space to initialize the initial position and initial velocity of each particle in the particle swarm algorithm, thereby obtaining the initial particle state vector set. Substitute the position vector of each particle in the initial particle state vector set into the stochastic optimization model for simulation evaluation, calculate the objective function value of maximizing expected return and the objective function value of risk at each state position, and determine the multi-objective fitness evaluation matrix that includes the two-dimensional objective value associated with each particle. Calculate the dominance relationships among the multi-objective fitness evaluation matrix, update the matrix based on the comparison of dominance relationships, and determine the individual historical optimal position and global optimal position of each particle. Based on the individual historical best position and the global best position, the position migration state of each particle after a set number of iterations in the multi-dimensional strategy space is continuously calculated to determine the Pareto optimal solution set. Based on the Pareto optimal solution set and the pre-introduced user multi-level risk preference coefficients, the target position vector is determined by the linear weighted matching method, and the target position vector is mapped to a set of day-to-day and real-time trading strategies that are robust under different risk preferences.

7. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 1, characterized in that, The transaction execution module includes: The strategy compilation module is configured to compile the received preferred strategy parameters into standardized trading instruction messages that conform to the communication protocol of the power trading platform; The contract execution module is configured to have dynamically configurable smart contract scripts deployed internally. The smart contract scripts encapsulate complete transaction logic for order placement, cancellation, and confirmation, and can monitor and judge multiple conditions such as preset price, time, trading volume, and price difference of related products in parallel based on real-time access to market data streams. The condition triggering module is configured to automatically invoke standardized trading instruction messages when market data is detected to meet the triggering conditions of any smart contract, execute the corresponding trading operation through the power trading platform API, and simultaneously generate an operation log with timestamps and execution environment hash values.

8. The power trading auxiliary decision-making system based on multi-source data and intelligent algorithms according to claim 1, characterized in that, The knowledge management module includes: The post-trade analysis module is configured to automatically perform the following analyses at three levels: daily, time-segment, and single transaction, based on transaction execution logs and actual market clearing data: Attribution analysis is performed on the deviation between the predicted values ​​and the actual values ​​output by the coupled prediction model; Compare the expected returns of the trading strategy with the actual risk-adjusted returns, and perform performance attribution; The knowledge graph update module is configured to structure and map the attribution conclusions, key performance indicators and corresponding market scenario characteristics obtained from the retrospective analysis module into entity nodes, attribute relationships and confidence labels in the knowledge graph. Among them, when the confidence level of the relationship between a specific scenario and a strategy pattern reaches a preset threshold, a model optimization instruction is triggered; Update the driver module to respond to model optimization commands, convert high-confidence knowledge graph triple data into an incremental learning sample set for coupled prediction models and constraint rules for policy optimization, and drive targeted iterative updates of coupled prediction models and policy rules.

9. A power trading auxiliary decision-making method based on multi-source data and intelligent algorithms, applied to the power trading auxiliary decision-making system based on multi-source data and intelligent algorithms as described in any one of claims 1-8, characterized in that, The method includes the following steps: Market, meteorological, and unit operation data are collected through a cloud-edge collaborative architecture, and after cleaning, verification, time alignment, and hash identification, they are stored in a distributed storage layer. The trained coupled prediction model is invoked to make synchronous predictions of future market factors. Combined with market rules and risk constraints inferred from the knowledge graph, a multi-objective optimization algorithm is used to generate multiple candidate trading strategies and conduct simulation comparisons to output a recommended strategy set. The recommended strategy is transformed into a smart contract script with trigger conditions, which is then deployed in the transaction execution module. When market conditions are met, the order placement operation is automatically executed, and the execution status and transaction results are monitored throughout the process. Based on actual transaction data and market results, the system performs automated retrospective analysis on prediction accuracy, strategy effectiveness, and execution compliance. The analysis results are then structured and fed back to a knowledge graph and coupled prediction model to enable the system to learn and continuously optimize itself.

10. The power trading auxiliary decision-making method based on multi-source data and intelligent algorithms according to claim 9, characterized in that, Based on actual transaction data and market results, an automated post-mortem analysis is conducted to assess forecast accuracy, strategy effectiveness, and execution compliance. This includes: Deviation analysis is performed between predicted and actual values ​​to identify sources of systematic error and adjust model features and structure. Compare the simulated returns of the strategy with the risk-adjusted returns after actual implementation to analyze whether the performance difference stems from forecast bias or unexpected market events. The new market patterns, effective strategy models, and lessons learned during this trading cycle will be constructed into new entities, relationships, and rules in a knowledge graph, and their confidence levels will be labeled. When the confidence level of new knowledge exceeds a set threshold, the incremental learning process of the coupled prediction model and the rule base update of the policy optimizer are triggered.

Images