A prediction method for power spot market price difference type discrimination

Abstract

The application discloses a kind of power spot market price difference type discrimination prediction method, it is related to power spot market prediction, electricity price difference analysis and risk identification technical field, comprising: collection meteorological satellite, numerical weather forecast, electricity price and market boundary data and pre-processing;Based on historical day-ahead-real-time price difference, construct the price difference type label system of numerical direction, fluctuation, cause and risk dimension;Extract market, statistical and mechanism characteristics and fusion;Variational mode decomposition is carried out to day-ahead electricity price, real-time electricity price and price difference sequence, the correlation of day-ahead electricity price and real-time electricity price is described in combination with t-Copula function, and the prediction value of price difference is obtained by adaptive fusion and correction using light gradient boosting machine model and time series prediction model based on self-attention mechanism;Then output certainty price difference type label and each type probability distribution.The application can reduce error accumulation, improve extreme price difference, special cause price difference and risk level discrimination precision.

Description

Technical Field

[0001] This invention relates to the fields of electricity spot market forecasting, electricity price spread analysis, and risk identification, and particularly to a forecasting method for identifying the type of electricity spot market price spread. Background Technology

[0002] The electricity spot market mainly comprises the day-ahead market and the real-time market. A day-ahead-real-time price spread is formed between the day-ahead market clearing price and the real-time market clearing price. This price spread is influenced by factors such as load fluctuations, uncertainty in renewable energy output, unit start-up and shutdown, grid congestion, and market participants' bidding strategies, exhibiting nonlinear, non-stationary, peak-tailed, and strongly time-series correlated characteristics.

[0003] Price spreads can typically be categorized based on their magnitude, volatility, direction, causal factors, and risk level. These include regular small price spreads, extreme positive price spreads, extreme negative price spreads, blocking price spreads, supply-demand imbalance price spreads, and abnormal volatility price spreads. Accurately identifying and predicting price spread types is crucial for electricity market participants in conducting spot market arbitrage transactions, risk management, unit combination optimization, and power purchase and sale decisions.

[0004] In existing technologies, most methods focus on predicting the specific value or direction of price spreads, without classifying and identifying the causes, volatility levels, and risk attributes of price spreads. This makes it difficult to meet the needs of market participants for price spread risk classification and strategy differentiation.

[0005] Meanwhile, existing technologies mostly use single models or simple combination models, failing to fully integrate the price difference formation mechanism to construct a multi-dimensional feature system. Some methods only use historical price differences, load and weather data, ignoring core features such as market bidding space, unit marginal cost, grid congestion level, new energy penetration rate and policy constraints, resulting in low accuracy in distinguishing extreme price differences and price differences with special causes.

[0006] In addition, some methods directly predict and subtract day-ahead and real-time electricity prices separately, or directly model the price difference series. The former ignores the strong correlation between the two types of prices, which can easily lead to amplified errors; the latter confuses the characteristics of different price fluctuations, making it difficult to accurately depict the inherent laws of price differences and has poor adaptability to scenarios of sudden changes in price difference types.

[0007] Some probabilistic prediction methods only output the probability distribution of price spreads, without integrating it with deterministic type discrimination. They cannot simultaneously provide the dual results of what type the price spread belongs to and the probability of that type occurring, thus limiting their practicality.

[0008] More importantly, price spread type identification is not simply a matter of numerical classification. Extreme positive price spreads, extreme negative price spreads, congestion-type price spreads, and supply-demand imbalance-type price spreads are often triggered by different market operating mechanisms. In the feature space, they manifest as a superposition of time-series fluctuations, market boundary constraints, new energy output deviations, and grid congestion. If post-processing classification is based solely on a single predicted value, it is difficult to explain the formation path of price spread types and to provide trading entities with actionable risk management guidelines.

[0009] Furthermore, the electricity spot market exhibits significant time continuity and scenario switching characteristics. The same market participant may face various states at different times, such as regular small price spreads, short-term peak price spreads, and persistent high-risk price spreads. If the model cannot place data preprocessing, labeling system, decomposition prediction, correlation correction, and probability output in the same computational chain, it is easy to encounter problems where prediction results, risk levels, and business interpretations are disconnected.

[0010] Therefore, there is a need for a method for identifying and predicting price spread types in the electricity spot market that can accurately identify price spread types, integrate probabilistic prediction and deterministic discrimination, deeply mine multi-dimensional features, and reduce error accumulation. Summary of the Invention

[0011] Technical problems to be solved

[0012] The purpose of this invention is to provide a prediction method for identifying the type of price difference in the electricity spot market, which solves the problems in the prior art such as missing price difference type identification, insufficient feature mining, improper handling of the correlation between day-ahead electricity price and real-time electricity price, error accumulation, and insufficient combination of probabilistic prediction results and deterministic type identification.

[0013] The core concept of this invention is to bring the price spread type labeling system forward to the training data construction stage and integrate this labeling system throughout the entire process of feature selection, decomposition prediction, correlation correction and result output, so that the model output no longer stops at a single price spread value, but forms a typified prediction result that can explain the source of risk.

[0014] Specifically, this invention does not simply string together multiple existing prediction models. Instead, it inputs the multi-scale decomposition results of day-ahead electricity prices, real-time electricity prices, and price spread sequences, along with market characteristics, statistical characteristics, and mechanistic characteristics, into the prediction chain. It also uses the joint distribution relationship between day-ahead electricity prices and real-time electricity prices to correct the preliminary prediction results, thereby forming a consistent data foundation between price spread prediction, type identification, and probability output.

[0015] Technical solution

[0016] To achieve the above objectives, the present invention provides the following technical solution: a prediction method for identifying the type of price difference in the electricity spot market, comprising the following steps:

[0017] Collect multi-source data related to the electricity spot market and preprocess the multi-source data to obtain a preprocessed multi-source dataset;

[0018] Based on the price difference between the historical previous market clearing price and the real-time market clearing price, a price difference type labeling system is constructed, and historical price difference samples are labeled according to the price difference type labeling system to obtain a labeled training dataset;

[0019] Based on the preprocessed multi-source dataset, market features, statistical features, and mechanistic features are extracted, and feature importance analysis and feature fusion are performed to obtain a fused feature set.

[0020] The variational mode decomposition model was used to decompose the preprocessed day-ahead electricity price series, real-time electricity price series and price difference series respectively, and the trend component, periodic component and random component of each series were obtained.

[0021] The t-Copula function is used to fit the joint distribution of day-ahead electricity price and real-time electricity price, quantify the nonlinear correlation between day-ahead electricity price and real-time electricity price, and obtain the correlation coefficient and the joint distribution residual deviation term;

[0022] The trend component and the periodic component are input into a time series prediction model based on a self-attention mechanism to extract long-term time series dependencies and obtain a first prediction result; the random component and the fused feature set are input into a lightweight gradient boosting machine model to mine the nonlinear relationships between features and obtain a second prediction result.

[0023] The contribution ratios of time-series features of the self-attention-based time-series prediction model and the contribution ratios of mechanistic nonlinear features of the lightweight gradient booster model are obtained respectively. The contribution ratios of time-series features and mechanistic nonlinear features are normalized to obtain the dual-model adaptive fusion weights. The first prediction result and the second prediction result are weighted and fused according to the dual-model adaptive fusion weights to obtain the preliminary price difference prediction result.

[0024] Based on the correlation coefficient and the joint distribution residual deviation term, the preliminary price difference prediction result is corrected for error to obtain the price difference prediction value;

[0025] Using the predicted price difference and the fused feature set as input, a multi-class logistic regression model is used to determine the price difference type based on the price difference type label system, outputting a deterministic price difference type label, and combining quantile regression to output the probability distribution of the predicted result belonging to each price difference type.

[0026] Preferably, after outputting the deterministic price difference type label and probability distribution, the day-ahead market clearing price after actual clearing, the real-time market clearing price and the verification results are written back to the sample database, and the data quality score, dimension contribution score, error correction coefficient and type confidence difference threshold are calibrated in a rolling manner to form a closed-loop link from data collection, model prediction, type identification to feedback correction.

[0027] Preferably, the multi-source data includes meteorological satellite data, numerical weather forecast data, electricity price data, and market boundary data. The electricity price data includes day-ahead market clearing prices and real-time market clearing prices. The market boundary data includes load, renewable energy output, non-market output, and inter-provincial transmission line data.

[0028] Preferably, the preprocessing includes standardizing the collected data, filling in missing values, removing outliers, and aligning various types of data according to timestamps to unify the granularity of the time scale.

[0029] Preferably, the price spread type labeling system includes a numerical direction dimension, a volatility dimension, a cause dimension, and a risk dimension; the numerical direction dimension includes extreme positive, normal positive, normal negative, and extreme negative; the volatility dimension includes low volatility, medium volatility, and high volatility; the cause dimension includes blocking type, supply and demand imbalance type, and abnormal volatility type; and the risk dimension includes low risk, medium risk, and high risk.

[0030] Preferably, the numerical direction dimension is divided after determining the threshold based on the historical price spread quantiles in the market, the volatility dimension is divided based on the price spread volatility, and the risk dimension is divided based on the price spread magnitude, volatility dimension, and price spread duration. Low risk is defined as a regular price spread with low volatility and a duration of less than 2 hours, medium risk is defined as a regular price spread with medium or high volatility and a duration of 2 to 4 hours, and high risk is defined as an extreme price spread with high volatility and a duration of more than 4 hours.

[0031] Preferably, the market characteristics include historical price spreads, day-ahead electricity prices, real-time electricity prices, total grid load, renewable energy output, non-market unit output, inter-provincial interconnection lines, meteorological satellite data, and numerical weather prediction data; the statistical characteristics include price spread volatility, skewness, kurtosis, load change rate, and temperature change rate; and the mechanistic characteristics include at least one of the following: bidding space, marginal cost, congestion surplus, renewable energy volatility coefficient, load peak-valley difference, and policy constraint factors.

[0032] Preferably, the variational mode decomposition model is used to decompose the day-ahead electricity price sequence, the real-time electricity price sequence, and the price difference sequence into multiple mode components, and to classify the multiple mode components into trend components, periodic components, and random components according to the frequency and fluctuation characteristics of the multiple mode components.

[0033] Preferably, the dual-model adaptive fusion weights include weights for a time-series prediction model based on a self-attention mechanism and weights for a lightweight gradient booster model. The weights for the time-series prediction model based on the self-attention mechanism are obtained by normalizing the contribution ratio of the time-series features, and the weights for the lightweight gradient booster model are obtained by normalizing the contribution ratio of the nonlinear features of the mechanism.

[0034] Preferably, the error correction includes: incorporating the correlation coefficient between the day-ahead electricity price and the real-time electricity price obtained by fitting the t-Copula function and the joint distribution residual deviation term into the preliminary price difference prediction result, so as to correct the systematic deviation caused by price coupling.

[0035] Preferably, the multi-class logistic regression model determines the numerical direction label, volatility label, cause label, and risk level label of the price spread in sequence according to the three-level discrimination process of coarse classification, fine discrimination, and risk rating, and outputs a complete deterministic price spread type label through the principle of maximum probability classification.

[0036] Beneficial effects

[0037] Compared with existing technologies, this invention provides a predictive method for identifying the type of price difference in the electricity spot market, which has the following beneficial effects:

[0038] 1. This invention constructs a price spread type labeling system and discrimination framework that includes numerical direction dimension, volatility dimension, cause dimension, and risk dimension, realizing multi-dimensional subdivision of price spread types. It can not only determine the positive or negative direction of price spreads, but also identify extreme price spreads, volatility types, cause types, and risk levels, and output deterministic labels, which facilitates market participants to formulate differentiated arbitrage and risk management strategies.

[0039] 2. This invention employs feature engineering that combines mechanism-driven and data-driven approaches, integrating the power market clearing mechanism, power grid operation patterns, and new energy characteristics to construct a fusion feature set that includes market characteristics, statistical characteristics, and mechanism characteristics. This differs from feature systems that only use basic meteorological and load data, and can improve the accuracy of identifying price differences caused by special factors.

[0040] 3. This invention employs a hybrid model combining decomposition, correlation, and fusion. It separates different fluctuation components through a variational mode decomposition model, characterizes the correlation between day-ahead electricity prices and real-time electricity prices through a t-Copula function, and fuses a lightweight gradient booster model with a time series prediction model based on a self-attention mechanism. This approach takes into account both global trends and local abrupt changes, effectively avoiding the problem of error amplification.

[0041] 4. This invention drives a multi-class logistic regression model by combining the price spread prediction value and the fusion feature set, so that the price spread prediction results can directly serve the classification of numerical direction, volatility, cause and risk level, avoiding the problem of only outputting numerical prediction results without being able to form a risk type conclusion.

[0042] 5. This invention outputs both probabilistic and deterministic results, providing both deterministic price spread type labels and the probability of occurrence for each type, offering market participants a dual decision-making basis combining risk quantification and type certainty. Furthermore, because this invention integrates the multi-dimensional labeling system, decomposition prediction model, and t-Copula function correlation correction, the predicted price spread, type labels, and probability distributions share a consistent data source and computational link, reducing decision-making bias caused by the separation of numerical prediction and type judgment.

[0043] 6. This invention can also provide an interpretable basis for subsequent trading strategies. For example, when the certainty label shows a high-risk price spread due to supply-demand imbalance and the probability distribution is concentrated in the high-risk type, market participants can adjust their power purchase and sale plans, risk exposure, or reserve capacity accordingly; when the probability distribution is dispersed, it can indicate the existence of type uncertainty in that period and reduce aggressive trading strategies. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall process of the prediction method for identifying the type of price difference in the electricity spot market according to the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the construction of the price difference type labeling system of the present invention.

[0046] Figure 3 This is a schematic diagram illustrating the construction of the "decomposition-correlation-fusion" model of this invention.

[0047] Figure 4 This is a schematic diagram illustrating the price difference type discrimination and probability-determinism dual output of the present invention. Detailed Implementation

[0048] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] Reference Appendix Figure 1 This embodiment provides a predictive method for identifying the type of price spread in the electricity spot market. The method first acquires multi-source data through data interfaces and devices, then preprocesses the multi-source data, constructs a price spread type labeling system, extracts and fuses features, constructs a "decomposition-correlation-fusion" model, and finally outputs the price spread type identification result and probability distribution result.

[0050] 1. Multi-source data acquisition and preprocessing

[0051] The system collects meteorological satellite data, numerical weather prediction data, electricity price data, and market boundary data. The electricity price data includes day-ahead market clearing prices and real-time market clearing prices. The market boundary data includes load, renewable energy output, non-market output, and inter-provincial power transmission lines.

[0052] After data collection, the data undergoes standardization, missing value imputation, outlier removal, and timestamp alignment to ensure uniform time scale granularity across all data types. This process creates a preprocessed multi-source dataset for subsequent processing.

[0053] In practical implementation, the timestamp alignment uses the clearing time granularity of the electricity spot market as a benchmark, mapping meteorological satellite data, numerical weather prediction data, total grid load, renewable energy output, non-market output, and inter-provincial tie-line data to the same time window. For data with a sampling period shorter than the clearing time granularity, aggregation can be performed within the corresponding time window; for data with a sampling period longer than the clearing time granularity, alignment can be performed according to the most recent valid time or forecast release time to ensure that various features are comparable at the same forecast time point.

[0054] It is also important to note that outlier removal does not simply involve deleting all outlier samples. Instead, it first performs business constraint verification by considering upper and lower limits of electricity prices, physical load boundaries, renewable energy installed capacity, and tie-line transmission capacity. For peak samples that statistically deviate from the mean but still meet power operation constraints, the system retains their sample records and adds a peak identifier, thereby avoiding the accidental deletion of key training samples needed to form extreme price differences.

[0055] For samples with missing values ​​imputed, the system generates a data quality identifier. This identifier is used to distinguish between truly missing data, delayed reporting, and imputed data after anomaly removal. It also serves as an auxiliary feature in the subsequent feature fusion process, enabling the model to identify the potential impact of data quality on the price difference type discrimination results.

[0056] The price difference between the current market clearing price and the real-time market clearing price is shown below:

[0057] ;

[0058] in, This represents the price difference between the market price before the corresponding time and the real-time market. The market clearing price for the corresponding time period; The real-time market clearing price at the corresponding moment; Used as time number.

[0059] Furthermore, to avoid implicit biases in price spread type discrimination caused by missing data completion or delayed reporting, the system can generate a data quality score based on the missing rate, delay rate, and anomaly correction rate after forming the preprocessed multi-source dataset. This score does not replace the original market features but serves as an auxiliary quality feature, participating in training along with the subsequent fused feature set, enabling the model to identify the impact of data source reliability on prediction results. The data quality score relationship is shown below:

[0060] ;

[0061] in, Assign a data quality score for the corresponding time point; , , These are the weights corresponding to the missing rate, latency rate, and outlier correction rate, respectively. , , These represent the proportion of missing data, the proportion of reporting delays, and the proportion of anomaly corrections within the same time window. This relationship allows for differentiated treatment of supplementary data and normally collected data during subsequent training, thereby improving the model's robustness to scenarios with fluctuating data quality.

[0062] 2. Construction of a Price Difference Type Labeling System

[0063] Reference Appendix Figure 2 Historical price spreads are calculated based on the day-ahead market clearing price and the real-time market clearing price, and a price spread type labeling system is constructed based on the historical price spread samples. The price spread type labeling system includes numerical direction dimension, volatility dimension, cause dimension, and risk dimension.

[0064] In terms of numerical direction, thresholds are determined by statistically analyzing historical market price spread quantiles. Based on these thresholds, price spread intervals are divided into four categories: extremely positive, normally positive, normally negative, and extremely negative. The specific criteria for this division are as follows:

[0065] when When, it is marked as extremely positive; when When, it is marked as normal positive; when When, it is marked as a normal negative; when When the value is 0, it is marked as extremely negative.

[0066] in, This is the positive extreme threshold determined based on the positive high quantile of historical price differences; This is the negative extreme threshold determined based on the negative lower quantile of historical price spreads.

[0067] In terms of volatility, price spreads are classified into three categories—low volatility, medium volatility, and high volatility—based on their volatility. Price spread volatility is obtained by calculating the dispersion of historical price spreads within a calculation window and is used to characterize the amplitude of price spread fluctuations over short periods.

[0068] In terms of causes, by combining information such as congestion surplus, inter-provincial tie-line constraints, load, renewable energy output, non-market output, price spread volatility, skewness, and kurtosis, price spread samples are labeled as congestion-type, supply-demand imbalance-type, or abnormal volatility-type. Among them, congestion-type price spreads are used to characterize price spreads caused by grid congestion or tie-line constraints; supply-demand imbalance-type price spreads are used to characterize supply-demand deviations caused by changes in load, renewable energy output, or non-market output; and abnormal volatility-type price spreads are used to characterize situations where price spreads change abruptly in a short period of time or deviate significantly from historical statistical characteristics.

[0069] In terms of risk, based on factors such as the magnitude and duration of price spreads, risk levels are divided into three categories: low risk, medium risk, and high risk. Low risk is defined as regular price spreads with low volatility and a duration of less than 2 hours; medium risk is defined as regular price spreads with medium or high volatility and a duration of 2 to 4 hours; and high risk is defined as extreme price spreads with high volatility and a duration of more than 4 hours. This classification rule can be flexibly adjusted according to the risk tolerance of market participants.

[0070] Using the above method, each historical price difference sample is labeled with a numerical direction label, a fluctuation label, a cause label, and a risk level label to form a labeled training dataset.

[0071] When labeling historical price difference samples, if the same sample simultaneously meets multiple causal dimension conditions, the primary causal label is determined according to the priority of grid congestion, supply-demand imbalance, and abnormal fluctuations, while the remaining causal factors are retained as auxiliary labels in the sample attributes. This approach ensures both the determinism of the training labels and the preservation of the complex causal information of price difference formation.

[0072] The price spread type labeling system is also used to constrain the output structure of the subsequent multi-class logistic regression model. Specifically, the numerical direction label is used to complete the coarse classification, the volatility label and the cause label are used to complete the fine discrimination, and the risk level label is used to complete the risk rating, so that the model prediction results can directly correspond to trading decisions and risk management scenarios.

[0073] For example, when the price spread of a historical sample is extremely positive, and there is a situation where the constraints of inter-provincial connecting lines are strengthened and the blockage surplus is significantly increased during the same period, the sample is labeled as extremely positive, corresponding to the volatility type, blockage type, and corresponding risk level; when the error in the prediction of new energy output expands and leads to a significant deviation in real-time supply and demand, the sample is labeled as supply and demand imbalance type. The above labeling logic enables the model to learn the correspondence between price spread type and market operating status.

[0074] 3. Feature Extraction and Fusion

[0075] Based on the preprocessed multi-source dataset, market features, statistical features, and mechanistic features are extracted.

[0076] Market characteristics include historical price differences, day-ahead electricity prices, real-time electricity prices, total grid load, renewable energy output, non-market unit output, inter-provincial interconnection lines, meteorological satellite data, and numerical weather prediction data.

[0077] Statistical characteristics include price spread volatility, skewness, kurtosis, load variability, and temperature variability, all calculated based on market characteristics. Skewness characterizes the degree of asymmetry in the price spread distribution, kurtosis characterizes the leptokurtic and thick-tailed characteristics of the price spread distribution, load variability characterizes short-term load variations, and temperature variability characterizes the impact of meteorological changes on load and electricity prices.

[0078] Mechanistic characteristics include bidding space, marginal cost, congestion surplus, renewable energy volatility coefficient, peak-valley load difference, and policy constraint factors. Bidding space characterizes the bidding or clearing space of market participants; marginal cost characterizes the change in marginal cost of generating units or the system; congestion surplus characterizes the impact of price differences caused by grid congestion; renewable energy volatility coefficient characterizes the uncertainty of renewable energy output; peak-valley load difference characterizes changes in system supply and demand pressure; and policy constraint factors characterize the impact of market rules or policy constraints on price formation.

[0079] Feature importance analysis is used to remove redundant features, retaining those that contribute significantly to spread type identification, resulting in a fused feature set. This fused feature set is used for subsequent lightweight gradient booster model prediction, spread type identification, and probability output.

[0080] In practice, feature importance analysis is performed in the following order: initial screening based on business mechanisms, evaluation of model contribution, and redundancy removal based on correlation. First, data fields that are clearly inconsistent with the formation mechanism of the electricity market or have been missing for a long time are removed. Second, effective features are selected based on the split gain, error reduction contribution, or ranking importance during model training. Finally, highly correlated features are merged or the one with the highest contribution is retained to avoid the model repeatedly learning the same market signal.

[0081] To enhance the model's ability to identify price spreads with special causes, the fusion feature set not only retains time-series statistics but also retains mechanistic features that characterize market operating constraints. For example, blocking surpluses and inter-provincial link constraints are used to support the identification of blocking price spreads, load peak-valley differences and new energy volatility coefficients are used to support the identification of supply-demand imbalance price spreads, and skewness and kurtosis are used to support the identification of abnormal volatility price spreads.

[0082] It is also important to note that feature fusion does not involve indiscriminately splicing together all features. Instead, it groups and maps features according to a price spread type labeling system. The numerical dimension focuses on the price spread magnitude, day-ahead electricity price, and real-time electricity price; the volatility dimension focuses on the price spread volatility, skewness, and kurtosis; the causal dimension focuses on the congestion surplus, renewable energy volatility coefficient, and load peak-valley difference; and the risk dimension focuses on the price spread duration, volatility level, and deviation from extreme thresholds.

[0083] In the feature fusion stage, to ensure a clear correspondence between the fused feature set and the price spread type labeling system, dimension contribution scores can be constructed for the numerical direction dimension, volatility dimension, causal dimension, and risk dimension. Each dimension contribution score is determined by the corresponding feature group and its feature contribution weights, used to select the feature subset most sensitive to that dimension label. The dimension contribution score relationship is shown below:

[0084]

[0085] in, Contribute a score to the features of a certain label dimension; This represents the summation range within the feature set belonging to this label dimension; The contribution weights of the corresponding features obtained during training; These are the corresponding normalized feature values. This scoring method allows features such as blocking surplus and inter-provincial link constraints to be preferentially mapped to the causal dimension, while price spread volatility, skewness, and kurtosis are preferentially mapped to the volatility dimension, thus avoiding the problem of unclear labeling criteria after feature fusion.

[0086] 4. Construction of the "Decomposition-Relevance-Fusion" Model

[0087] Reference Appendix Figure 3 A "decomposition-correlation-fusion" model is constructed. The model includes multi-scale decomposition, correlation modeling, hybrid prediction model, and error correction.

[0088] 4.1 Multiscale Decomposition

[0089] Variational mode decomposition (VMD) was used to decompose the preprocessed daytime electricity price series, real-time electricity price series, and price spread series, respectively, to obtain the trend component, periodic component, and stochastic component of each series. The VMD relationships of the series are shown below:

[0090] ;

[0091] in, For the first The value of the sequence at the corresponding time; Assign a sequence category number; Number the modal components; For the first Class sequence number The values ​​of each modal component at the corresponding time; This indicates the summation of all modal components.

[0092] Multiple modal components obtained from variational mode decomposition are merged based on frequency and fluctuation characteristics. Low-frequency and smoothly changing modal components are merged into trend components; modal components with stable periodic fluctuation characteristics are merged into periodic components; and high-frequency and highly abrupt modal components are merged into random components. Trend components are used to characterize the long-term trend of the sequence, periodic components are used to characterize the periodic changes of the sequence, and random components are used to characterize random perturbations and local mutations in the sequence.

[0093] Variational mode decomposition can separate different volatility characteristics at three levels: day-ahead electricity price, real-time electricity price, and price spread, avoiding prediction bias caused by the mixture of trend, periodic and random fluctuations when directly modeling the original sequence.

[0094] When determining the modal component merging results, the system combines the center frequency, energy proportion, and the magnitude of change in adjacent time periods. Low-frequency modal components with relatively stable energy proportions are classified as trend components, modal components with intraday or weekly periodic fluctuations are classified as periodic components, and high-frequency modal components with significant energy abrupt changes and high correlation with extreme price difference samples are classified as random components.

[0095] This invention decomposes the day-ahead electricity price series, real-time electricity price series, and price spread series simultaneously, rather than modeling only the single price spread series. This is because day-ahead and real-time electricity prices are formed under different market mechanisms, and their trend components and sources of random disturbances are not entirely consistent. By decomposing the three types of series separately, it is possible to extract the combined characteristics of price spread changes while preserving the price's own volatility characteristics.

[0096] During the model training phase, the trend component, cyclical component, and stochastic component each carry different market information: the trend component reflects long-term supply and demand structure changes, the cyclical component reflects intraday load peaks and troughs and trading cycle patterns, and the stochastic component reflects short-term disturbances such as new energy output deviations, unplanned unit conditions, or sudden blockages. This component division provides the input basis for the subsequent division of labor between the two models.

[0097] 4.2 Correlation Modeling

[0098] Since the relationship between day-ahead and real-time electricity prices is not a simple linear one, and the correlation may be enhanced under extreme price differences or abnormal fluctuations, this invention employs the t-Copula function, a connection function based on the t-distribution, to fit the joint distribution of day-ahead and real-time electricity prices. This quantifies the nonlinear correlation between day-ahead and real-time electricity prices and calculates the correlation coefficient and the joint distribution residual deviation term. The t-Copula function is shown below:

[0099] ;

[0100] in, For degrees of freedom and correlation parameters The t-Copula function; The degrees of freedom are t-distributed. This is a correlation parameter between day-ahead electricity prices and real-time electricity prices; The probability of the marginal distribution of day-ahead electricity prices; This represents the marginal distribution probability of real-time electricity prices.

[0101] The correlation modeling described above can provide a basis for error correction in subsequent price spread predictions, avoiding the error amplification problem caused by single variable modeling or subtracting the day-ahead and real-time electricity prices after independent prediction.

[0102] In practice, the correlation coefficient output by the t-Copula function reflects the strength of the linkage between day-ahead and real-time electricity prices within the current training window, while the joint distribution residual deviation term reflects the degree of deviation between the joint distribution of the two and the actual observed samples. When extreme price difference samples appear in clusters, this residual deviation term can provide a directional basis for error correction.

[0103] By incorporating the correlation modeling results into error correction, this invention does not simply subtract the two electricity price forecasts, but instead uses the joint distribution relationship between day-ahead electricity prices and real-time electricity prices to constrain the preliminary price difference forecast results, thereby reducing the systematic bias caused when the price coupling relationship is ignored.

[0104] It is also important to note that the t-Copula function is not used as the sole output of the probabilistic prediction model in this invention. Instead, it serves as an intermediate tool for characterizing the correlation between day-ahead and real-time electricity prices and for error correction. This setting avoids the separation between the probabilistic model and the deterministic discriminant model, allowing the correlation modeling results to directly serve the final price spread prediction and type determination.

[0105] 4.3 Hybrid Prediction Model

[0106] The decomposed trend and periodic components are input into a time series prediction model based on a self-attention mechanism to extract long-term temporal dependencies and obtain the first prediction result. This time series prediction model is suitable for modeling long-term temporal dependencies in trend and periodic components, thereby obtaining global trend and periodic fluctuation information.

[0107] The decomposed random components and the fused feature set are input into a lightweight gradient booster model to mine the nonlinear relationships between features and obtain a second prediction result. The lightweight gradient booster model is suitable for handling the nonlinear relationships between market features, statistical features, and mechanistic features in the fused feature set, and is particularly suitable for identifying local mutations and special causal changes reflected in the random components.

[0108] During model training, the time-series prediction model based on the self-attention mechanism focuses on learning the dependencies across time steps in the trend and periodic components, while the lightweight gradient boosting machine model focuses on learning the nonlinear mapping relationship between the stochastic component and market boundaries, statistical features, and mechanistic features. These two models correspond to two different levels of prediction information: long-term time-series patterns and local mutation mechanisms.

[0109] To avoid the failure of a single model under different market conditions, the prediction contribution of the two models can be calculated separately for regular price difference samples, extreme price difference samples, and special cause price difference samples during the training phase, and the adaptive fusion weights for different sample scenarios can be determined accordingly.

[0110] For example, when price spread changes are mainly driven by intraday load cycles and day-ahead market trends, the contribution of time series forecasting models based on self-attention mechanisms increases; when price spread changes are mainly driven by blockage surplus mutations, new energy output deviations, or changes in tie-line constraints, the contribution of lightweight gradient booster models increases.

[0111] The contribution ratios of temporal features from the self-attention-based temporal prediction model and the contribution ratios of mechanistic nonlinear features from the lightweight gradient boosting machine model are obtained separately, and both are normalized to obtain the adaptive fusion weights for the two models. The adaptive fusion rules for the two models are as follows:

[0112] ;

[0113] in, This is a preliminary forecast of the price spread; Weights for a time-series prediction model based on a self-attention mechanism; This is the first prediction result output by the time series prediction model based on the self-attention mechanism; For the weights of the lightweight gradient booster model; This is the second prediction result output by the lightweight gradient booster model; The proportion of contributions to time series features; The proportion of contribution to the nonlinear characteristics of the mechanism.

[0114] Through the above weighted fusion, preliminary price spread prediction results are obtained, achieving deep integration of time series trend prediction results and nonlinear mutation prediction results.

[0115] When trend and cyclical components contribute significantly to price spread changes, the weight of the time-series prediction model based on the self-attention mechanism increases accordingly; conversely, when stochastic components and mechanistic features contribute significantly to price spread changes, the weight of the lightweight gradient booster model increases accordingly. Thus, the fusion model can automatically adjust its dependence on time-series patterns and abrupt changes based on market conditions.

[0116] 4.4 Error Correction

[0117] Based on the preliminary price spread forecast results, the nonlinear correlation coefficient between the day-ahead electricity price and the real-time electricity price obtained by fitting the t-Copula function and the joint distribution residual deviation term are introduced to correct the error in the weighted fusion preliminary price spread forecast results. The error correction rules are as follows:

[0118] ;

[0119] in, This is the corrected price spread forecast; This is a preliminary forecast of the price spread; These are the error correction coefficients; The residual deviation term is determined by the joint distribution residuals of the t-Copula function.

[0120] Error correction can eliminate the superposition of errors caused by independent forecasting of two types of electricity prices, correct the systematic deviation caused by price coupling, and finally output a high-precision and robust price difference forecast.

[0121] In actual operation, error correction is performed using a rolling window. After each new trading session forecast is completed, the day-ahead electricity price, real-time electricity price, and price difference sample after actual clearing are written back to the correction sample library, and the correlation coefficient and joint distribution residual deviation term are updated, so that subsequent forecasts can continuously adapt to changes in market structure.

[0122] If the proportion of extreme price difference samples increases within the rolling window, the weight of the joint distribution residual bias term in error correction is increased; if the price difference remains within a normal, small fluctuation range within the rolling window, the degree of disturbance of the residual bias term to the preliminary prediction results is reduced. This approach balances prediction stability in normal scenarios with responsiveness in extreme scenarios.

[0123] During the error correction phase, to avoid over-correction or under-correction under different market conditions using fixed error correction coefficients, the system can dynamically adjust the error correction coefficients based on the residual dispersion within the rolling window. The dynamic error correction coefficient relationship is shown below:

[0124] ;

[0125] in, This refers to the dynamic error correction coefficient used at the current moment; This is the benchmark correction strength under normal market conditions; This is an adjustment factor used to control the effect of residual dispersion on the correction strength; This represents the volatility of the joint distribution residuals within the most recent forecast window. The correction strength increases as the number of extreme price difference samples increases and the residual dispersion rises; when the market is stable, the correction strength remains near the baseline level, thus balancing forecast stability and responsiveness to extreme scenarios.

[0126] 5. Price Spread Type Identification and Result Output

[0127] Reference Appendix Figure 4 Using the predicted price spread and the fused feature set as input, a multi-class logistic regression model is constructed. Based on a pre-defined price spread type labeling system, the multi-class logistic regression model sequentially identifies the numerical direction type, volatility type, cause type, and risk level of the price spread, and determines the labels for each dimension through the maximum probability classification principle, ultimately outputting complete deterministic price spread type labels.

[0128] Specifically, first, a coarse classification label for the numerical direction dimension is determined based on the predicted price spread; then, fine discriminative labels for the volatility and causal dimensions are determined by combining the fused feature set; finally, a risk level label is determined based on the price spread magnitude, volatility type, and duration. This results in a deterministic price spread type label that includes numerical direction labels, volatility labels, causal labels, and risk level labels.

[0129] Simultaneously, quantile regression is used to output the probability of occurrence for each price spread type. The quantile regression takes the predicted price spread value and the fused feature set as input, and outputs the price spread prediction result or type probability estimate at different quantile levels, thereby forming the probability distribution of the predicted result belonging to each price spread type. Therefore, this invention can simultaneously output deterministic price spread type labels and the probability distribution of the predicted result belonging to each price spread type, forming a dual output of probabilistic and deterministic results.

[0130] When outputting results, if the type label of the certainty spread matches the highest probability type in the probability distribution, then that type is output as a high confidence level; if the two do not match, or if the difference between the highest probability and the second highest probability is less than the preset confidence threshold, then the prediction period is marked as a sample to be reviewed, and market participants are prompted to reduce their trading exposure or adopt a conservative risk strategy.

[0131] To clarify the relationship between deterministic labels and probability distributions, the system can also calculate the type confidence difference. The type confidence difference measures the degree of separation between the highest probability type and the second highest probability type, and serves as the basis for triggering a review prompt. The type confidence difference relationship is shown below:

[0132] ;

[0133] in, Type confidence difference for the current forecast period; This is the highest probability of occurrence among all price spread types. This represents the probability of occurrence of the second highest probability type, excluding the highest probability type. When the type confidence difference is greater than the preset confidence difference threshold, it indicates that the deterministic spread type label and the probability distribution have high consistency; when the type confidence difference is less than or equal to the preset confidence difference threshold, it indicates that there is a competitive relationship between multiple spread types, and the system marks the prediction period as a low-confidence output and prompts the adoption of a conservative trading strategy.

[0134] In an optional embodiment, the system receives subsequent actual clearing data and manual review results as feedback signals. If the feedback result indicates a misclassification of a certain type of label, the corresponding sample and its fused features are added to the retraining sample set, and the numerical direction threshold, volatility threshold, or risk level rules are rolled over for calibration. If the feedback result confirms that the model prediction is correct, the effective weight of similar samples in subsequent training is increased. This closed-loop mechanism improves the model's adaptability to new market volatility scenarios.

[0135] In a specific application, the root mean square error of this method can be reduced to 12.3 yuan / MWh, which is more than 25% lower than the average of 16.5 yuan / MWh of the existing technology.

[0136] Example 2:

[0137] In another embodiment, the present invention can also be deployed in a transaction support decision-making platform for electricity market participants. The transaction support decision-making platform receives the price spread prediction value, deterministic price spread type label, and probability distribution of each type output by the method, and generates corresponding prompt information based on different price spread types.

[0138] When the output result is an extreme positive price difference and the probability distribution is concentrated in the high-risk type, the system prompts market participants to pay attention to the risk of rising real-time market procurement costs; when the output result is an extreme negative price difference and the duration exceeds the preset time, the system prompts market participants to pay attention to the risk of deviation in electricity sales revenue; when the output result is a congestion-type price difference, the system prioritizes displaying correlation characteristics such as inter-provincial tie line constraints, congestion surplus, and regional load changes, so that market participants can judge the cause of the price difference.

[0139] In this implementation, the deterministic price spread type label is used to quickly provide a trading decision direction, the probability distribution is used to measure the confidence level of this decision direction, and the key contributing features in the fusion feature set are used to provide auxiliary explanations. Through the above configuration, this invention can not only output prediction conclusions but also provide market operation evidence supporting these conclusions, thereby improving the usability of algorithmic prediction results in actual trading scenarios.

[0140] In summary, this invention achieves accurate identification and prediction of price spread types in the electricity spot market through multi-source data acquisition and preprocessing, construction of a price spread type labeling system, feature extraction and fusion, variational mode decomposition, t-Copula function correlation modeling, adaptive fusion prediction of a lightweight gradient booster model and a time series prediction model based on a self-attention mechanism, and quantile regression probability output.

[0141] 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, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A predictive method for identifying the type of price difference in the electricity spot market, characterized in that, Includes the following steps: Step 1: Collect multi-source data related to the electricity spot market and preprocess the multi-source data to obtain a preprocessed multi-source dataset; Step 2: Based on the price difference between the historical previous market clearing price and the real-time market clearing price, construct a price difference type labeling system, and label the historical price difference samples according to the price difference type labeling system to obtain a labeled training dataset; Step 3: Based on the preprocessed multi-source dataset, extract market features, statistical features, and mechanistic features, and perform feature importance analysis and feature fusion to obtain a fused feature set; Step 4: Use variational mode decomposition model to decompose the preprocessed day-ahead electricity price series, real-time electricity price series and price spread series respectively to obtain the trend component, periodic component and random component of each series; Step 5: Use the t-Copula function to fit the joint distribution of day-ahead electricity price and real-time electricity price, quantify the nonlinear correlation between day-ahead electricity price and real-time electricity price, and obtain the correlation coefficient and the joint distribution residual deviation term; Step 6: Input the trend component and periodic component into the time series prediction model based on the self-attention mechanism, extract long-term time series dependencies and obtain the first prediction result; The random components and the fused feature set are input into a lightweight gradient booster model to mine the nonlinear relationships between features and obtain a second prediction result. Step 7: Obtain the contribution ratio of the time series features of the self-attention mechanism-based time series prediction model and the contribution ratio of the mechanistic nonlinear features of the lightweight gradient booster model, respectively. Normalize the contribution ratio of the time series features and the contribution ratio of the mechanistic nonlinear features to obtain the dual-model adaptive fusion weights. Then, perform weighted fusion of the first prediction result and the second prediction result according to the dual-model adaptive fusion weights to obtain the preliminary price difference prediction result. Step 8: Based on the correlation coefficient and the joint distribution residual deviation term, perform error correction on the preliminary price difference prediction result to obtain the price difference prediction value; Step 9: Using the predicted price difference and the fused feature set as input, a multi-class logistic regression model is used to determine the price difference type based on the price difference type label system, outputting a deterministic price difference type label, and combining quantile regression to output the probability distribution of the predicted result belonging to each price difference type.

2. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step one, the multi-source data includes meteorological satellite data, numerical weather forecast data, electricity price data, and market boundary data. The electricity price data includes day-ahead market clearing prices and real-time market clearing prices. The market boundary data includes load, renewable energy output, non-market output, and inter-provincial transmission line data.

3. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step one, the preprocessing includes standardizing the collected data, filling in missing values, removing outliers, aligning various types of data according to timestamps to unify the granularity of the time scale, and generating data quality identifiers to distinguish between true missing data, delayed reporting, and outlier removal.

4. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step two, the price spread type labeling system includes numerical direction dimension, volatility dimension, cause dimension, and risk dimension; the numerical direction dimension includes extreme positive, normal positive, normal negative, and extreme negative; the volatility dimension includes low volatility, medium volatility, and high volatility; the cause dimension includes blocking type, supply and demand imbalance type, and abnormal volatility type; the risk dimension includes low risk, medium risk, and high risk.

5. The prediction method for identifying the type of price difference in the electricity spot market according to claim 4, characterized in that, The numerical direction dimension is divided according to the threshold determined by the historical price spread quantile of the market. The volatility dimension is divided according to the price spread volatility. The risk dimension is divided according to the price spread magnitude, volatility dimension and price spread duration. Low risk is a regular price spread with low volatility and a duration of less than 2 hours. Medium risk is a regular price spread with medium or high volatility and a duration of 2 to 4 hours. High risk is an extreme price spread with high volatility and a duration of more than 4 hours.

6. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step three, the market characteristics include historical price spreads, day-ahead electricity prices, real-time electricity prices, total grid load, renewable energy output, non-market unit output, inter-provincial interconnection lines, meteorological satellite data, and numerical weather prediction data; the statistical characteristics include price spread volatility, skewness, kurtosis, load change rate, and temperature change rate; and the mechanistic characteristics include at least one of the following: bidding space, marginal cost, congestion surplus, renewable energy volatility coefficient, load peak-valley difference, and policy constraint factors.

7. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step four, the variational mode decomposition model is used to decompose the day-ahead electricity price series, real-time electricity price series and price difference series into multiple modal components, and to classify the multiple modal components into trend components, periodic components and random components according to the center frequency, energy ratio and fluctuation change characteristics of the multiple modal components.

8. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step seven, the dual-model adaptive fusion weights include the weights of the time-series prediction model based on the self-attention mechanism and the weights of the lightweight gradient booster model. The weights of the time-series prediction model based on the self-attention mechanism are obtained by normalizing the contribution ratio of the time-series features, and the weights of the lightweight gradient booster model are obtained by normalizing the contribution ratio of the nonlinear features of the mechanism.

9. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step eight, the error correction includes: incorporating the correlation coefficient and joint distribution residual deviation term between the day-ahead electricity price and the real-time electricity price obtained by fitting the t-Copula function into the preliminary price difference prediction result, and updating the correlation coefficient and joint distribution residual deviation term according to the rolling window to correct the systematic deviation caused by price coupling.

10. The prediction method for identifying the type of price difference in the electricity spot market according to claim 1, characterized in that, In step nine, the multi-class logistic regression model follows a three-level discrimination process of coarse classification, fine discrimination, and risk rating. First, it determines the numerical direction label based on the predicted price spread value. Then, it determines the volatility label and the causal label based on the volatility characteristics and causal characteristics in the fusion feature set. Finally, it determines the risk level label based on the price spread amplitude, volatility label, and duration, and outputs a complete deterministic price spread type label through the maximum probability classification principle.

Images