A method for forecasting power spot market volume

Abstract

The present application relates to the technical field of power transaction, in particular to a kind of power spot transaction volume prediction method;It includes the following steps: multi-source data acquisition: obtain the historical transaction data of power market, real-time market data, weather data, user side response data and power grid real-time state data, the historical transaction data includes historical transaction volume, historical electricity price, market participant bidding information, the real-time market data includes the market supply and demand state of current period, transmission line capacity, the weather data includes temperature, humidity, wind speed and extreme weather warning information, the user side response data includes user electricity behavior mode, demand elasticity coefficient, the power grid real-time state data includes node load rate, equipment operating state;The present application can effectively solve the problem of large deviation of traditional method prediction, significantly improve the accuracy and market adaptability of power spot transaction volume prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transaction, in particular to a power spot transaction volume prediction method. BACKGROUND

[0002] In the field of power market spot transaction, accurately predicting transaction volume is the key to ensuring market stable operation and improving resource allocation efficiency. Existing power spot transaction volume prediction techniques have significant defects: most of them only rely on historical transaction data and simple meteorological parameters, and predict through a single machine learning model, which is difficult to fully capture the complex dynamic characteristics of the power market. In actual transactions, the power market is influenced by the interaction of multiple factors such as real-time operation state of power grid, user electricity behavior change, and intermittent output of new energy. For example, the operation constraints such as power grid transmission line congestion and node overload will directly limit the actual occurrence of power transaction, but the existing methods do not consider such key constraints; at the same time, the influence of dynamic factors such as user demand elasticity and market participant strategy game on transaction volume cannot be accurately quantified due to the traditional fixed weight allocation mechanism. This leads to a large deviation between the prediction results and the actual transaction situation, which cannot provide reliable transaction decision basis for market participants, and seriously restricts the scientific management and efficient operation of the power market.

[0003] Based on the above problems, there is an urgent need for a power spot transaction volume prediction technology scheme that can integrate multi-source dynamic data, quantify power grid operation constraints, and adaptively adjust data weights to improve prediction accuracy and market adaptability. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art, and to propose a power spot transaction volume prediction method, comprising the following steps:

[0005] Multi-source data acquisition: obtaining historical transaction data, real-time market data, meteorological data, user side response data and real-time state data of power grid of the power market,

[0006] The historical transaction data includes historical transaction volume, historical electricity price, market participant bidding information, the real-time market data includes market supply and demand state, transmission line capacity of the current period, the meteorological data includes temperature, humidity, wind speed and extreme weather warning information, the user side response data includes user electricity behavior mode, demand elasticity coefficient, and the real-time state data of power grid includes node load rate, equipment operation state;

[0007] Feature engineering construction: standardizing the collected multi-source data, selecting features through an improved hybrid model of discrete particle swarm optimization algorithm and support vector regression, extracting a feature subset strongly related to power spot transaction volume, the feature subset includes meteorological influence factor, user elasticity coefficient, market price fluctuation index and power grid load correction term;

[0008] Hybrid model training: a hybrid prediction model containing a long short-term memory network and a graph neural network is constructed, wherein the long short-term memory network is used to capture the time sequence characteristics in the time sequence data, and the graph neural network is used to model the interaction relationship between market participants, the model parameters are optimized by a bidirectional propagation algorithm to form an initial prediction model;

[0009] Real-time correction and prediction: real-time current market data is collected, the power grid load correction term is calculated and input into the initial prediction model, the influence weight of each data source is adjusted by combining a dynamic weight distribution mechanism, the future T period power spot transaction volume prediction result is output, and the dynamic weight distribution mechanism adjusts the weight coefficients of meteorological data, user side response data and power grid state data according to market state and time window.

[0010] Preferably, in the multi-source data collection step, the meteorological data is obtained by fusing meteorological satellites, ground monitoring stations and numerical weather prediction models, and the user side response data is constructed by intelligent electric meters, user electricity behavior questionnaire survey and demand response event history record; the historical transaction data is obtained from the power trading center database, the real-time market data is real-time grabbed through the power market information interaction platform, and the power grid real-time state data is collected through the power dispatching automation system.

[0011] Further preferably, in the feature engineering construction step, the standardization processing includes Z-score standardization for numerical data and one-hot encoding for categorical data; the improved discrete particle swarm optimization algorithm introduces a differential evolution operator and a dynamic inertia weight adjustment strategy to enhance global search ability and avoid falling into local optimum; when performing feature selection, the prediction error minimization is taken as the objective function, and the cross-validation method is used to evaluate the prediction performance of different feature combinations.

[0012] Further preferably, in the hybrid model training step, the graph neural network adopts an attention mechanism to capture the dynamic interaction relationship between market participants, and the weight calculation of the attention mechanism is based on the historical transaction behavior, market share and quote strategy similarity of the participants; the long short-term memory network sets multiple hidden layers to prevent overfitting through Dropout technology; during the model training process, an adaptive learning rate adjustment algorithm is used to dynamically adjust the learning rate according to the training error.

[0013] Further preferably, in the real-time correction and prediction step, the power grid load correction term is calculated by the following formula:

[0014] ;

[0015] wherein, is the power grid load correction coefficient, is the current node load rate, is the node rated load rate, Real-time power flow for transmission lines, Due to the capacity limitation of transmission lines, and is the weight parameter obtained by training based on historical data.

[0016] Further preferably, the dynamic weight allocation mechanism is implemented by the following formula:

[0017] ;

[0018] in, For the Class data source in The weight of the moment, For the Pearson correlation coefficient between class data source and transaction volume, To control the adjustment parameter of weight sensitivity, is the total number of data source categories.

[0019] Further preferably, the output result of the hybrid prediction model is corrected by the following formula:

[0020] ;

[0021] in, For the The revised forecast trading volume at the time, is the number of features in the feature subset, For the The LSTM or GNN model output corresponding to the feature, is the grid load correction factor, For the Dynamic weights of features.

[0022] Further preferably, a model updating step is included: regularly collecting new market data, and using an online learning algorithm to perform incremental training on the hybrid prediction model, wherein the online learning algorithm includes stochastic gradient descent and adaptive moment estimation to update the model parameters in real time; during the model updating process, an update threshold is set, and when the difference between the new data and the historical data exceeds the threshold, the model update is triggered.

[0023] Further preferably, the method also includes a prediction result evaluation step: using root mean square error, mean absolute percentage error and Sharpe ratio to perform multi-dimensional evaluation of the prediction results, and the Sharpe ratio is used to measure the risk-adjusted return of the prediction results; based on the evaluation results, the prediction model is optimized and adjusted, and when the evaluation indicators do not meet the preset standards, feature selection and model training are re-performed.

[0024] Further preferably, the method is applied to an electricity market transaction decision support system, the system comprising a data acquisition module, a feature engineering module, a model training module, a real-time correction module and a result visualization module, the modules interact with each other through a message queue to realize low-delay prediction in a high-concurrency scenario.

[0025] Technical effects: The present application makes technical breakthroughs through multi-source data fusion, dynamic weight distribution and hybrid model construction, aiming at the problem that traditional technologies do not comprehensively consider power grid operation constraints and market dynamic change factors. On the one hand, the data of power grid load rate, power transmission line power flow, etc. are included in the prediction system, and the influence of power grid constraints is quantified through correction formula; on the other hand, the weight is dynamically adjusted based on data correlation, and the time sequence and market subject interaction features are mined by combining the advantages of LSTM and GNN. Thus, the problem of large prediction deviation of traditional methods is effectively solved, and the accuracy and market adaptability of electricity spot transaction volume prediction are significantly improved, providing a reliable basis for electricity market transaction decision. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The flowchart of the electricity spot transaction volume prediction method of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0028] Please refer to Figure 1 The traditional electricity spot transaction volume prediction technology has the following technical problems: it relies too much on a single data source or limited types of data, making it difficult to fully cover complex factors affecting transaction volume; feature selection lacks an efficient optimization mechanism, resulting in redundant model inputs or missing key features; a single prediction model is used, which cannot take into account both time sequence features and market subject interaction relationships; and there is a lack of real-time correction mechanism, making it difficult to respond to market dynamic changes.

[0029] Based on this, the embodiment provides a power spot transaction volume prediction method, including the following steps: a multi-source data acquisition step acquires historical transaction data, real-time market data, meteorological data, user-side response data and real-time grid state data of the power market, and detailedly defines specific information covered by each type of data; a feature engineering construction step standardizes the collected multi-source data, uses a hybrid model of improved discrete particle swarm optimization algorithm (MDPSO) and support vector regression (SVR) for feature selection, and extracts a feature subset strongly related to the power spot transaction volume; a hybrid model training step constructs a hybrid prediction model containing long short-term memory network (LSTM) and graph neural network (GNN), respectively gives play to the advantages of LSTM in capturing time sequence features and GNN in modeling the interaction relationship of market participants, and optimizes model parameters through a bidirectional propagation algorithm; a real-time correction and prediction step acquires current market data in real time, calculates a grid load correction term and inputs the initial prediction model, adjusts the influence weight of each data source in combination with a dynamic weight distribution mechanism, and outputs a power spot transaction volume prediction result of a future T period.

[0030] The scheme compensates for the defect of single data dimension of traditional methods through comprehensive collection of multi-source data, ensures the integrity of input information, effectively filters out key features through a hybrid algorithm for feature selection, improves the training efficiency and accuracy of the model, realizes double modeling of time sequence and market subject relationship through the combination of LSTM and GNN, and breaks through the limitation of a single model, and the real-time correction and dynamic weight adjustment mechanism can quickly respond to market changes and ensure that the prediction result fits the actual situation.

[0031] The technical effects achieved by the above embodiment include: ensuring the input quality of the prediction model from the data source, laying a foundation for accurate prediction; an optimized feature selection process avoids redundant data interference, so that the model focuses on core influencing factors; a hybrid model architecture deeply mines market data features and accurately describes the transaction volume change law; a real-time correction mechanism gives the model dynamic adaptability, significantly improves the timeliness and reliability of the prediction, and provides strong support for power market transaction decision-making.

[0032] The traditional power data acquisition has the following technical problems: the data acquisition channels are scattered and lack systematic integration, resulting in insufficient data integrity; some key data sources are single, and the data accuracy and timeliness are difficult to guarantee; different types of data collection methods are not standardized, affecting subsequent data processing and analysis.

[0033] Based on this, in the multi-source data collection step, meteorological data is obtained by fusing meteorological satellites, ground monitoring stations and numerical weather prediction models, integrating space-based, ground-based and model prediction data to ensure the comprehensiveness and accuracy of meteorological information; user-side response data is constructed by intelligent electric meters, user electricity behavior questionnaire surveys and demand response event history records, multi-dimensionally collecting user electricity information from real-time electricity data, subjective behavior preferences to historical response events, fully presenting user-side characteristics; historical transaction data is obtained from the electric power trading center database to ensure the authority and integrity of the data; real-time market data is real-time scraped through the electric power market information interaction platform to realize the instant update of the data; real-time grid state data is collected through the electric power dispatching automation system (SCADA) to ensure the accuracy and real-time of the grid operation data.

[0034] This scheme solves the problems of dispersion, singleness and non-standardization of traditional data collection by clearly defining the accurate collection channels and methods of various types of data. The fusion acquisition method of meteorological data avoids the information loss caused by the weather monitoring blind area of a single data source and can fully capture the influence of weather changes on electricity demand; the multi-channel construction of user-side response data breaks the limitation of relying only on smart meter data and deeply excavates user behavior characteristics; the standardized data collection source guarantees the quality and reliability of the data, providing a solid data foundation for subsequent feature engineering and model training.

[0035] The technical effects achieved by the above embodiments include: a complete, accurate and real-time multi-source data collection system is constructed to reduce data errors and omissions; high-quality input data is provided for the electric power spot transaction volume prediction model to improve the model's representation ability for complex market factors; the data can truly reflect the operation state of the electric power market and the change of user demand, enhancing the credibility and practicality of the prediction results.

[0036] The traditional feature engineering has the following technical problems: the data standardization processing method is single and cannot effectively adapt to the characteristics of different types of data; the feature selection algorithm is easy to fall into local optimum and it is difficult to find a global optimal feature combination; there is a lack of scientific evaluation mechanism and it is difficult to accurately judge the influence of feature combination on prediction performance.

[0037] Therefore, in the feature engineering construction step, the Z-score standardization is performed on the numerical data to eliminate the influence of data dimension and make the data conform to the standard normal distribution, so as to facilitate model processing; the one-hot encoding is performed on the discrete type data to convert the category information into a vector form recognizable by the machine. The improved MDPSO algorithm introduces the differential evolution operator and the dynamic inertia weight adjustment strategy. The differential evolution operator enhances the global search ability of the algorithm, and the dynamic inertia weight adjusts the search step according to the iteration process to avoid the algorithm falling into local optimum. In the feature selection, the minimum prediction error is taken as the objective function, the cross-validation is used to evaluate the prediction performance of different feature combinations, and the generalization ability of the feature combination is comprehensively evaluated by dividing the training set and the test set multiple times.

[0038] The scheme adapts to the characteristics of different data types through diversified standardization processing, improves the effectiveness of data processing, and effectively overcomes the defects of the traditional algorithm through the improved MDPSO algorithm, which can select the most representative feature subset from a large number of features. The evaluation mechanism based on cross-validation scientifically judges the advantages and disadvantages of the feature combination and avoids the performance degradation of the model caused by improper feature selection.

[0039] The technical effects achieved by the above embodiments include: an optimized data preprocessing process improves data quality and model training efficiency; an accurate feature selection process reduces redundant feature interference and reduces model complexity; and the optimal feature combination found through scientific evaluation significantly improves the prediction accuracy and generalization ability of the model, so that the model can stably and accurately predict the power spot transaction volume in different scenarios.

[0040] The traditional model training has the following technical problems: a single neural network model is difficult to simultaneously process time series features and complex market subject relationships; the model structure design is unreasonable, which is prone to overfitting phenomenon; and the learning rate is fixed during the training process, which cannot be dynamically adjusted according to the training situation, affecting the convergence speed and performance of the model.

[0041] Therefore, in the hybrid model training step, the combination architecture of GNN and LSTM is adopted. The GNN adopts an attention mechanism to capture the dynamic interaction relationship between market participants, calculates the attention weight based on the historical transaction behavior of the participants, market share and pricing strategy similarity, accurately describes the game and cooperation relationship between market subjects, and the LSTM sets multiple hidden layers to increase the extraction ability of the model to time series data features, and prevents overfitting phenomenon through the Dropout technology to randomly discard part of the neurons.

[0042] In the model training process, an adaptive learning rate adjustment algorithm is adopted to dynamically adjust the learning rate according to the training error. In the early stage of training, a higher learning rate is used to speed up the convergence, and when approaching the optimal solution, the learning rate is reduced to improve the convergence accuracy.

[0043] This scheme solves the limitations of traditional single models through the complementary advantages of GNN and LSTM. LSTM focuses on mining long-term dependencies in time series, while GNN focuses on modeling the interaction between market participants. The combination of the two fully captures the characteristics of power market data. Reasonable model structure design and the application of Dropout technology effectively avoid overfitting and improve the generalization ability of the model. The adaptive learning rate adjustment algorithm makes the model training more efficient and can quickly find the optimal parameters.

[0044] The technical effects achieved by the above embodiments include: the constructed hybrid model can deeply mine the time series rules and subject interaction patterns in the power market data, accurately predict the transaction volume trend; the optimized model structure and training strategy improve the stability and reliability of the model, reduce the prediction deviation caused by data fluctuations; the adaptive learning rate speeds up the model training, reduces the training cost, and enables the model to quickly adapt to market data changes and update the prediction ability in a timely manner.

[0045] The traditional power spot transaction volume prediction has the following technical problems when considering the influence of grid state: ignoring the influence of actual operation constraints of the grid on the transaction volume, leading to the disconnection between the prediction results and the actual market transaction; lacking an effective method to quantify the influence of grid load on transaction volume, unable to accurately correct the prediction results.

[0046] Therefore, in the real-time correction and prediction step, the grid load correction term is calculated by the formula:

[0047] ;

[0048] , where is the grid load correction coefficient, is the current node load rate, is the node rated load rate, is the real-time power flow of the transmission line, is the transmission line capacity limit, and are weight parameters trained according to historical data.

[0049] In the power spot transaction volume prediction, the actual operation state of the grid has an important influence on the transaction, but the traditional method often ignores this key factor. The formula is used to calculate the grid load correction coefficient , which aims to quantify the influence of the grid operation state on the prediction results of the transaction volume, making the prediction more consistent with the actual transaction. In the formula, The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid. The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid.

[0050] The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid.

[0051] The node load rate represents the current node load rate, which reflects the actual load degree of each node in the power system at a certain time, and is a key indicator for measuring the local operation pressure of the power grid.

[0052] Traditional data weight allocation has the following technical problems: the fixed weight allocation method cannot adapt to the dynamic changes of the power market and it is difficult to reflect the differences in the importance of different data sources under different market conditions; the lack of scientific weight calculation methods leads to unreasonable data weight settings, affecting the accuracy of prediction results.

[0053] Based on this, this embodiment provides a dynamic weight allocation mechanism through the formula:

[0054] ;

[0055] in For the Class data source in The weight of the moment, For the Pearson correlation coefficient between class data source and transaction volume, To control the adjustment parameter of weight sensitivity, The total number of data source categories. This formula calculates weights based on the correlation between data sources and trading volume, amplifies correlation differences through an exponential function, and uses normalization to ensure that the total weight is 1, achieving dynamic and adaptive weight allocation.

[0056] In the electricity market, the impact of different data sources on spot electricity trading volume varies with market conditions and time. Traditional fixed weight allocation methods cannot adapt to this dynamic nature. This formula implements a dynamic weight allocation mechanism that adjusts the weight of each data source in the forecast in real time based on its correlation with trading volume, thereby optimizing the forecast model's adaptability to market changes.

[0057] In the formula, Indicates the Class data source in The weight at a given moment is a quantitative reflection of the influence of the data source on the prediction result at the current moment. Representative Class data source in Time data, such as It can be the first Real-time meteorological data, user-side response data, etc. For the The actual electricity spot trading volume at the moment.

[0058] It is The Pearson correlation coefficient between a data source and trading volume measures the degree of linear correlation between the two. Its value ranges from -1 to 1, with the closer the absolute value is to 1, the stronger the correlation. This coefficient is calculated by calculating the covariance and standard deviation of the data source and trading volume data, and objectively reflects the impact of the data source on trading volume.

[0059] The adjustment parameter controls the sensitivity of the weight, which amplifies or reduces the impact of correlation differences on the weight. When the value is large, the weight of the data source with high correlation will be significantly increased, while the weight of the data source with low correlation will be greatly reduced, so that the model pays more attention to the key data source; on the contrary, when the value is small, the weight adjustment is relatively flat, and the weight difference of each data source is reduced.

[0060] The denominator normalizes the numerator to ensure that the sum of all weights is 1, ensuring the rationality and effectiveness of weight allocation. Through this dynamic weight calculation method based on correlation, in different market scenarios such as extreme weather, holidays, etc., the weight of the data source closely related to the trading volume is automatically increased, thereby improving the response ability and prediction accuracy of the prediction model to market dynamic changes.

[0061] This scheme breaks the limitations of traditional fixed weights through a dynamic weight allocation mechanism, which can adjust the weight of each data source in real time according to market conditions and time windows. In different market scenarios such as extreme weather, major holidays, etc., the weight of the data source with high correlation to trading volume (such as weather data, user-side response data) is automatically increased, highlighting its importance to prediction. The technical effects achieved by the above embodiments include: enabling the prediction model to dynamically adapt to the complex and variable environment of the electricity market, improving the model's response ability to different market conditions; a reasonable weight calculation method ensures that each data source weight matches the actual influence, optimizing data utilization efficiency and improving prediction accuracy; through dynamic weight adjustment, the model's adaptability to market abnormal fluctuations and unexpected events is enhanced, making the prediction result more reliable and practical.

[0062] The traditional electricity spot trading volume prediction has the following technical problems: lack of comprehensive correction mechanism for prediction results, unable to effectively integrate the adjustment of power grid load impact and multi-source data weight change on prediction results; the prediction result does not fully consider the complex constraints and dynamic factors in the actual market operation, resulting in large prediction deviation.

[0063] Based on this, the output result of the hybrid prediction model is corrected by the formula:

[0064]

[0065] , where is the corrected prediction trading volume at the moment, is the number of features in the feature subset, is the prediction trading volume before correction at the moment.the LSTM or GNN model output corresponding to the i-th feature, is the grid load correction coefficient, is the dynamic weight of the i-th feature.

[0066] Traditional power spot trading volume prediction lacks comprehensive consideration of the impact of grid load and dynamic changes in data weights, resulting in deviations between predicted results and actual situations.

[0067] The formula is used to correct the output results of the hybrid prediction model, integrating the grid load correction term and the dynamic weight distribution result, so that the prediction is more consistent with the actual operation of the power market. In the formula, is the corrected prediction trading volume at time t, which is the final output prediction result.

[0068] is the grid load correction coefficient described above, which reflects the impact of real-time operation of the grid on trading volume, is used to adjust the basic prediction result, when the grid load is high, the value increases, the predicted trading volume is adjusted accordingly, reflecting the limiting effect of grid constraints on trading.

[0069] is the number of features in the feature subset, i.e. the number of key features selected after feature engineering for input into the prediction model. is the dynamic weight of the i-th feature at time t, calculated by the dynamic weight distribution mechanism formula, which determines the importance of the feature in the current time prediction. is the LSTM or GNN model output corresponding to the i-th feature,

[0070] is the LSTM or GNN model output corresponding to the i-th feature, the LSTM model is good at processing time series data and capturing the time series variation of trading volume; the GNN model focuses on modeling the interaction between market participants and mining market dynamic features. The outputs of the two are weighted and summed by dynamic weights to obtain the basic prediction value

[0071] The formula combines the grid load correction term with the dynamic weight distribution result to comprehensively adjust the prediction result from two key dimensions: grid operation status and data importance, fully considering the complex factors affecting power spot trading volume, significantly improving the accuracy and reliability of the prediction, providing more valuable trading volume prediction information for power market participants, and helping them to develop scientific and reasonable trading strategies.

[0072] ​​​​​The formula combines the power grid load correction term with the dynamic weight distribution result to comprehensively correct the output of the basic prediction model, and comprehensively considers the influence of the power grid operation state and the data weight change on the transaction volume. The scheme builds a perfect prediction result adjustment system through the comprehensive correction formula, and makes up for the defects of the traditional prediction lacking a correction mechanism. The dynamic adjustment of the power grid load and the data weight into the prediction result calculation makes the prediction more in line with the actual market operation. The technical effects achieved by the above embodiments include: significantly improving the accuracy of the power spot transaction volume prediction, reducing the prediction error caused by not considering the power grid constraints and the dynamic change of data; through comprehensive correction, the processing capacity of the prediction model for complex market factors is enhanced, and the prediction result is closer to the actual transaction situation; more reliable transaction volume prediction information is provided for power market participants to assist them in formulating scientific and reasonable transaction strategies, and the efficiency of the power market operation and the rationality of resource allocation are improved.

[0073] The traditional model updating has the following technical problems: the offline training method cannot respond to market data changes in time, leading to a decline in model prediction ability over time; there is no effective mechanism to trigger model updating, making it difficult to determine when to optimize the model, and the model may be too old or over-updated.

[0074] Therefore, the method further includes a model updating step: periodically collecting new market data, and using an online learning algorithm to incrementally train the hybrid prediction model, the online learning algorithm including stochastic gradient descent (SGD) and adaptive moment estimation (Adam) to update the model parameters in real time; during the model updating process, an update threshold is set, and when the difference between the new data and the historical data exceeds the threshold, the model updating is triggered. This step realizes the continuous optimization of the model through online learning, and quickly updates the parameters using an efficient optimization algorithm; the setting of the update threshold provides a scientific update decision basis, balancing the timeliness and stability of model updating.

[0075] The scheme solves the problems of offline training and blind updating of the traditional model through online learning and update threshold mechanism. Online learning enables the model to continuously absorb new data information and maintain adaptability to market changes; the update threshold avoids the instability caused by frequent updates of the model due to small data fluctuations, while ensuring timely optimization when there are significant changes in the market.

[0076] The technical effects achieved by the above embodiments include: significantly improving the long-term prediction performance of the model, enabling it to accurately predict the power spot transaction volume continuously; the scientific update mechanism reduces the maintenance cost of the model and avoids invalid or excessive training; the model's adaptability to the dynamic evolution of the power market is enhanced, and it can still maintain high prediction accuracy in changing scenarios such as market rule adjustment and new energy access, providing stable and reliable prediction support for power market transactions.

[0077] The traditional prediction result evaluation has the following technical problems: the evaluation index is single, only focuses on the prediction error, and cannot comprehensively measure the quality of the prediction result; there is a lack of model optimization closed loop based on the evaluation result, so that the evaluation result cannot be effectively fed back to the model improvement, and it is difficult to continuously improve the prediction performance.

[0078] Therefore, the method further comprises a prediction result evaluation step: multi-dimensional evaluation of the prediction result is performed by using root mean square error (RMSE), mean absolute percentage error (MAPE) and Sharpe ratio; the prediction model is optimized and adjusted according to the evaluation result, and when the evaluation index does not reach the preset standard, feature selection and model training are performed again.

[0079] This step analyzes the prediction result from different angles by using multiple evaluation indexes, the RMSE measures the overall deviation of the predicted value and the true value, the MAPE reflects the relative error, and the Sharpe ratio evaluates the risk and return; based on the optimization adjustment mechanism of the evaluation result, a complete closed loop from evaluation to improvement is formed.

[0080] The scheme solves the limitations of the traditional evaluation through multi-dimensional evaluation and closed loop optimization mechanism. Multiple evaluation indexes comprehensively cover the accuracy, relative error size and risk and return characteristics of the prediction result, avoiding the one-sidedness of a single index; based on the model optimization of the evaluation result, the model can be continuously improved according to the actual performance, and the prediction performance is continuously improved.

[0081] The technical effects achieved by the above embodiments include: comprehensive and scientific evaluation of the prediction result is realized, and accurate basis is provided for model optimization; through the closed loop optimization mechanism, the model is continuously improved, and the prediction accuracy and stability are improved; the power market participants can more accurately understand the quality of the prediction result, make reasonable trading decisions based on reliable evaluation, and continuously evolve and improve the prediction model in actual application.

[0082] The traditional power market prediction application has the following technical problems: there is a lack of integrated application system, data processing, model training and prediction result display are independent of each other, resulting in low work efficiency; the system architecture cannot meet the real-time requirements in the high concurrency scenario, and it is difficult to process a large amount of data and output the prediction result in a short time; the prediction result display form is single, and it is not convenient for users to intuitively understand and analyze.

[0083] Based on this, the method is applied to a power market transaction decision support system, the system includes a data acquisition module, a feature engineering module, a model training module, a real-time correction module and a result visualization module, the modules interact with each other through a message queue to realize low-delay prediction in a high-concurrency scenario; the result visualization module uses a variety of chart forms such as heat maps and line charts to intuitively display the prediction results and their influencing factors. The system integrates the whole process of power spot transaction volume prediction through modular design, and the message queue realizes efficient data transmission and processing; diversified visualization forms intuitively present complex data.

[0084] The scheme solves the problems of efficiency, real-time performance and display of traditional applications by constructing an integrated decision support system. The modular architecture and message queue technology ensure stable operation of the system in a high-concurrency scenario, and quickly complete data processing and prediction; the rich visualization forms facilitate users to quickly obtain key information and deeply analyze the prediction results and influencing factors.

[0085] The technical effects achieved by the above embodiments include: significantly improving the application efficiency of power market transaction volume prediction, shortening the time from data acquisition to result output; meeting the real-time transaction decision-making needs of the power market, providing timely and accurate prediction information for users; intuitive visualization helps users understand market trends and prediction results more clearly, assists them in developing scientific and reasonable transaction strategies, and improves the intelligent and scientific level of power market transactions.

[0086] The above is only a preferred embodiment of the present application, and does not limit the form of the present application, any skilled person in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the technical solution of the present application.

Claims

1. A method for forecasting power spot market volume, characterized by, The method comprises the following steps: Multi-source data collection: obtaining historical transaction data, real-time market data, weather data, user-side response data, and real-time grid state data of the electricity market, The historical transaction data includes historical transaction volume, historical electricity price, and market participant bidding information, the real-time market data includes market supply and demand status and transmission line capacity of the current period, the weather data includes temperature, humidity, wind speed, and extreme weather warning information, the user-side response data includes user electricity consumption behavior pattern and demand elasticity coefficient, and the real-time grid state data includes node load rate and equipment operating state; Feature engineering construction: standardizing the collected multi-source data, selecting features by using a hybrid model of an improved discrete particle swarm optimization algorithm and a support vector regression, and extracting a feature subset strongly related to electricity spot trading volume, wherein the feature subset includes weather influencing factors, user elasticity coefficient, market price fluctuation index, and grid load correction term; Hybrid model training: constructing a hybrid prediction model containing a long short-term memory network and a graph neural network, wherein the long short-term memory network is used to capture time sequence features in time series data, the graph neural network is used to model the interaction relationship between market participants, model parameters are optimized by a bidirectional propagation algorithm, and an initial prediction model is formed; Real-time correction and prediction: collecting current market data in real time, calculating the grid load correction term and inputting the initial prediction model, adjusting the influence weight of each data source in combination with a dynamic weight distribution mechanism, and outputting electricity spot trading volume prediction results of the future T time period, wherein the dynamic weight distribution mechanism adaptively adjusts the weight coefficients of weather data, user-side response data, and grid state data according to market state and time window; In the real-time correction and prediction step, the grid load correction term is calculated by the following formula: ； wherein, is the grid load correction factor, is the current node load rate, is the node rated load rate, is the real-time power flow of the transmission line, is the transmission line capacity limit, and is the weight parameter trained according to historical data; The dynamic weight distribution mechanism is realized by the following formula: ； wherein, is the weight of the jth data source at the ith time point, is the weight of the jth data source at the ith time point, is the weight of the jth data source at the ith time point, is the Pearson correlation coefficient between the jth data source and the trading volume at the ith time point, is the Pearson correlation coefficient between the jth data source and the trading volume at the ith time point, is the adjustment parameter for controlling the sensitivity of the weight, is the total number of data source categories; The output result of the hybrid prediction model is corrected by the following formula: ； wherein, is the is the is the number of features of the feature subset, is the is the LSTM or GNN model output corresponding to the is the grid load correction coefficient, is the is the weight of the is the weight of the data source of the 2. The method of claim 1, wherein, In the multi-source data collection step, the weather data is obtained by fusing meteorological satellites, ground monitoring stations, and numerical weather prediction models, and the user-side response data is constructed by using intelligent electric meters, user electricity consumption behavior questionnaire surveys, and demand response event historical records; the historical transaction data is obtained from the electricity trading center database, the real-time market data is real-time grabbed through the electricity market information interaction platform, and the real-time grid state data is collected through the power dispatching automation system.

3. The method of claim 1, wherein, In the feature engineering construction step, the standardization processing includes Z-score standardization for numerical data and one-hot encoding for categorical data; the improved discrete particle swarm optimization algorithm introduces a differential evolution operator and a dynamic inertia weight adjustment strategy to enhance global search ability and avoid falling into local optimum; when performing feature selection, the minimum prediction error is taken as the objective function, and the prediction performance of different feature combinations is evaluated in a cross-validation manner.

4. The method of claim 1, wherein, In the hybrid model training step, the graph neural network adopts an attention mechanism to capture the dynamic interaction relationship between market participants, and the weight calculation of the attention mechanism is based on the historical trading behavior, market share and quote strategy similarity of the participants; the long short-term memory network sets multiple hidden layers and prevents overfitting through the Dropout technology; in the model training process, an adaptive learning rate adjustment algorithm is adopted to dynamically adjust the learning rate according to the training error.

5. The method of forecasting the volume of power spot transactions according to claim 1, wherein The method also includes a model updating step: periodically collecting new market data, and incrementally training the hybrid prediction model using an online learning algorithm, which includes stochastic gradient descent and adaptive M-estimation, to update the model parameters in real time; in the model updating process, an updating threshold is set, and when the difference between the new data and the historical data exceeds the threshold, the model updating is triggered.

6. The method of forecasting the volume of power spot transactions according to claim 1, wherein The method also includes a prediction result evaluation step: using the root mean square error, the mean absolute percentage error and the Sharpe ratio to evaluate the prediction results in multiple dimensions, and the Sharpe ratio is used to measure the risk-adjusted returns of the prediction results; according to the evaluation results, the prediction model is optimized and adjusted, and when the evaluation indicators do not reach the preset standard, the feature selection and model training are performed again.

7. The method of forecasting the volume of power spot transactions according to claim 1, characterized in that, The method is applied to a power market transaction decision support system, and the system includes a data acquisition module, a feature engineering module, a model training module, a real-time correction module and a result visualization module, and the modules interact with each other through a message queue to realize low-latency prediction in a high-concurrency scenario.

Images