Probabilistic prediction method for clearing price of electricity market

Abstract

The invention relates to the technical field of electricity market analysis and prediction, in particular to a probabilistic prediction method for the clearing price of an electricity market. The technical problems are that when an electricity market clearing price probabilistic prediction method in the prior art is practically applied, multi-peak distribution caused by market state structural change is difficult to accurately capture, a predicted price probability density curve is often too smooth, and a prediction result has systematic deviation; according to the technical scheme, the probabilistic prediction method for the clearing price of the electricity market comprises a market state mode division step, a mode probability prediction step, a condition price distribution prediction step and a probability distribution synthesis step. According to the method, recognition and division of the market state mode are introduced, and a complex probability prediction problem is decomposed into a plurality of simple sub-problems, so that the multi-peak characteristic of the clearing price probability distribution is more accurately captured, and the prediction precision is improved.

Description

Technical Field

[0001] This invention relates to the field of electricity market analysis and forecasting technology, and in particular to a probabilistic forecasting method for electricity market clearing prices. Background Technology

[0002] As the penetration rate of intermittent and fluctuating renewable energy sources such as wind power and solar power continues to increase in the electricity market, the uncertainty of power system operation has significantly increased. This uncertainty is directly transmitted to the electricity market, making the fluctuations in market clearing prices more drastic and complex. Accurate probabilistic forecasting (i.e., predicting the probability of future prices falling within different ranges) is crucial for market participants' bidding strategies, risk assessments, and the safe and stable operation of the power grid.

[0003] Existing probabilistic prediction methods for electricity market clearing prices can be mainly divided into two categories: one is based on statistical models (such as quantile regression, GARCH model, etc.), and the other is based on machine learning models (such as LSTM neural network, random forest, etc.) combined with technologies such as Dropout or Bootstrap. These methods usually take the uncertainty of renewable energy output as the main input and obtain the probability distribution of prices through a large number of scenario simulations or direct regression.

[0004] However, in practical applications, most of these existing technologies treat the uncertainty of renewable energy output and the bidding game behavior of market participants as independent factors, and use end-to-end black box mapping for modeling. This approach ignores the physical mechanism of marginal unit switching in the formation of electricity market clearing prices. When the net load (total load minus renewable energy output) changes significantly, the marginal units in the market (i.e., the units that determine the market clearing price) will switch identities, such as switching from cheap coal-fired power units to expensive gas-fired power units, or even zero or negative electricity prices. This switching leads to a strong nonlinear, non-smooth, and even abrupt relationship between prices and influencing factors (such as net load), which is characterized by a multi-peak distribution in probability distribution.

[0005] Existing probabilistic forecasting methods struggle to accurately capture the multi-peaked distribution resulting from structural changes in market conditions. The price probability density curves they predict are often too smooth, exhibiting a single peak shape. This severely underestimates the probability of the market being in extreme states (such as high or negative prices), leading to systematic biases in the forecast results. For market participants, this bias can directly trigger incorrect bidding decisions and uncontrollable risk exposure.

[0006] Therefore, to address the above problems, a probabilistic prediction method for electricity market clearing prices is proposed. By introducing the identification and classification of market state patterns, the complex probabilistic prediction problem is decomposed into multiple simple sub-problems, thereby more accurately capturing the multi-peak characteristics of the clearing price probability distribution and improving prediction accuracy. Summary of the Invention

[0007] To overcome the problem that existing probabilistic prediction methods for electricity market clearing prices have difficulty accurately capturing the multi-peaked distribution caused by structural changes in market conditions, and that the predicted price probability density curves are often too smooth, leading to systematic biases in the prediction results.

[0008] The technical solution of this invention is: a probabilistic prediction method for electricity market clearing prices, comprising the following steps: Market state pattern classification steps: Based on historical market operation data, extract multi-dimensional features related to the clearing price formation mechanism, and use unsupervised learning algorithms to classify historical market states into a finite number of discrete price formation patterns; Pattern probability prediction step: Based on the results of the market state pattern classification step, construct and train a pattern probability prediction model to output the probability of belonging to each price formation pattern in a specific future period; Conditional price distribution prediction steps: For each price formation pattern identified, construct and train a conditional probability prediction model to output the conditional probability distribution of the market clearing price given that the pattern occurs. Probability distribution synthesis step: For a specific future period to be predicted, the probability of each model output by the model probability prediction step is combined with the corresponding conditional probability distribution output by the conditional price distribution prediction step, and the final probability distribution of the market clearing price for that period is calculated by weighted synthesis.

[0009] Preferably, the market state pattern segmentation step extracts multidimensional features closely related to the clearing price formation mechanism based on historical market operation data; these multidimensional features are analyzed using unsupervised learning algorithms to automatically divide the historically complex and continuous market operation state into a finite number of discrete "price formation patterns" with clear physical meaning; this step discretizes the continuous nonlinear problem of price formation through state recognition preprocessing; the pattern probability prediction step constructs and trains a pattern probability prediction model based on the pattern results from the previous step; this model can output the probability of belonging to each price formation pattern for a specific future period based on the predicted information; this step realizes the prediction of the possible structural state of the future market; the conditional price distribution prediction step targets the segmentation... Each price formation pattern is selected from historical data, and all samples belonging to that pattern are identified, with a conditional probability prediction model provided for each. Since the market mechanism is relatively stable within the same pattern, the relationship between price and influencing factors becomes smoother and simpler. Therefore, these sub-models can learn and predict the conditional probability distribution of the clearing price under the specific pattern's occurrence. In the probability distribution synthesis step, for the specific future period to be predicted, the probability of each pattern output from the pattern probability prediction step is used as a weight and multiplied by the corresponding conditional probability distribution output from the conditional price distribution prediction step. Then, all weighted distributions are superimposed to synthesize the final probability distribution of the market clearing price for that period. This step integrates the possibilities under different market conditions using the total probability formula.

[0010] Preferably, in the market state pattern segmentation step, the multidimensional features related to the clearing price formation mechanism include at least the following: Characteristic quantities that represent the net load level of a system are used to directly reflect the basic supply and demand balance of the system. Characteristic quantities that represent the pricing behavior of key pre-defined generating units are used to capture the core game behavior of market participants; Characteristic quantities that represent the overall supply and demand slack of the system, used to provide more capacity margin information belonging to net load; By integrating these multidimensional characteristics, the market conditions that determine the clearing price can be reflected.

[0011] Preferably, the characteristic quantity representing the net load level of the system is composed of the difference between the load forecast value and the renewable energy output forecast value, and further includes its statistical distribution characteristics. The feature quantity characterizing the bidding behavior of the preset key units is used to filter the key unit set based on the frequency of occurrence of marginal units in historical clearing data, and the bidding information of the units in the set is extracted. The characteristic quantity that characterizes the overall supply and demand balance of the system is defined by the ratio of the difference between the total available generating capacity and the net load forecast to the net load forecast. The larger the ratio, the more relaxed the system is.

[0012] Preferably, the unsupervised learning algorithm used in the market state pattern segmentation step is the Gaussian mixture model clustering algorithm; the Gaussian mixture model can fit the complex probability distribution of features, thereby discovering naturally formed clusters in the data; Each price formation pattern obtained through clustering corresponds to a market operating state dominated by a specific type of marginal unit, giving the identified patterns a clear physical and economic explanation.

[0013] Preferably, the pattern probability prediction model constructed in the pattern probability prediction step is a multi-class probability prediction model. The input features of this model are predictive features for a specific future period. These predictive features are isomorphic to the multidimensional features used in the market state pattern segmentation step, ensuring the consistency of information transmission. The output of the model is the probability value of each price formation pattern belonging to a specific future time period. The probability value is a probability vector.

[0014] Preferably, the multi-class probability prediction model is implemented using a Softmax regression model or a gradient boosting decision tree model; both of these models can provide good classification probability calibration.

[0015] Preferably, in the conditional price distribution prediction step, for any price formation pattern, the training data of its conditional probability prediction model is strictly derived from a subset of historical data that has been classified into that pattern. The conditional probability prediction model is implemented using a quantile regression model or a quantile regression forest model. These two models can directly output the conditional quantiles of variables, thus conveniently constructing a complete conditional probability distribution.

[0016] Preferably, for any price formation mode, the input features of its conditional probability prediction model are a simplified feature set that is strongly correlated with the clearing price under that mode. The simplified feature set includes at least the net load level and the price of key units. This simplification avoids overfitting and highlights the core influencing factors under that mode.

[0017] Preferably, the weighted synthesis method in the probability distribution synthesis step is as follows: The conditional probability distribution corresponding to each price formation pattern is multiplied by the probability of occurrence of that pattern obtained through the pattern probability prediction model. The weighted probability distributions of all patterns are then superimposed to form the final probability distribution.

[0018] Preferably, the method is specifically applied to probabilistic price forecasting in a day-ahead clearing scenario of the electricity market containing a high proportion of intermittent renewable energy; in this scenario, the output of renewable energy is highly uncertain and marginal unit switching is frequent.

[0019] The beneficial effects of this invention are: In this invention, the pattern partitioning step decomposes the complex nonlinear global problem; the pattern probability prediction step enables forward-looking judgment of future market conditions; the conditional distribution prediction step achieves high-precision learning within relatively simple linear subspaces; and the final synthesis step fully restores the overall uncertainty of the market. This not only improves prediction accuracy, especially the ability to capture tail risks in prices, but also endows the model with interpretability through the pattern concept, enabling decision-makers to clearly understand the driving factors behind price fluctuations and make reasonable decisions. This achieves more accurate and reliable probabilistic predictions of electricity market clearing prices, especially in environments with a high proportion of renewable energy. This invention uses the structural state of market operation as the pivot of modeling, and through a discretization-conditionalization-resynthesis path, it solves the problem that traditional models struggle to reflect the multi-peak characteristics of price probability distributions. Attached Figure Description

[0020] Figure 1 The diagram shown is a simplified flowchart of the probabilistic prediction method for electricity market clearing prices according to the present invention. Figure 2 The diagram illustrates a detailed process of the probabilistic prediction method for electricity market clearing prices according to the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Example 1 Please see Figure 1 and Figure 2 This invention provides an embodiment: a probabilistic prediction method for electricity market clearing prices, comprising the following steps: Market state pattern classification steps: Based on historical market operation data, extract multi-dimensional features related to the clearing price formation mechanism, and use unsupervised learning algorithms to classify historical market states into a finite number of discrete price formation patterns; Pattern probability prediction steps: Based on the results of the market state pattern segmentation steps, construct and train a pattern probability prediction model to output the probability of belonging to each price formation pattern in a specific future period; Conditional price distribution prediction steps: For each price formation pattern identified, construct and train a conditional probability prediction model to output the conditional probability distribution of the market clearing price given that the pattern occurs. Probability distribution synthesis step: For a specific future period to be predicted, the probability of each model output by the comprehensive model probability prediction step and the corresponding conditional probability distribution output by the conditional price distribution prediction step are combined in a weighted synthesis method to calculate the final probability distribution of the market clearing price for that period.

[0023] The process includes several steps. First, a market state pattern segmentation step extracts multidimensional features closely related to the clearing price formation mechanism based on historical market data. Unsupervised learning algorithms are then used to analyze these features, automatically dividing the historically complex and continuous market states into a finite number of discrete "price formation patterns" with clear physical meaning. This step preprocesses the continuous nonlinear problem of price formation through state recognition. Next, a pattern probability prediction step constructs and trains a pattern probability prediction model based on the pattern results from the previous step. This model can output the probability of a specific price formation pattern for a given future period, based on predicted information. This step enables the prediction of the possible structural state of the future market. Finally, a conditional price distribution prediction step addresses the segmentation... Each price formation pattern is selected from historical data, and all samples belonging to that pattern are identified, with a conditional probability prediction model provided for each. Since the market mechanism is relatively stable within the same pattern, the relationship between price and influencing factors becomes smoother and simpler. Therefore, these sub-models can learn and predict the conditional probability distribution of the clearing price under the specific pattern's occurrence. In the probability distribution synthesis step, for the specific future period to be predicted, the probability of each pattern output from the pattern probability prediction step is used as a weight and multiplied by the corresponding conditional probability distribution output from the conditional price distribution prediction step. Then, all weighted distributions are superimposed to synthesize the final probability distribution of the market clearing price for that period. This step integrates the possibilities under different market conditions using the total probability formula.

[0024] Furthermore, in the market state pattern segmentation steps, the multidimensional characteristics related to the clearing price formation mechanism include at least the following: Characteristic quantities that represent the net load level of a system are used to directly reflect the basic supply and demand balance of the system. Characteristic quantities that represent the pricing behavior of key pre-defined generating units are used to capture the core game behavior of market participants; Characteristic quantities that represent the overall supply and demand slack of the system, used to provide more capacity margin information belonging to net load; By integrating these multidimensional characteristics, the market conditions that determine the clearing price can be reflected.

[0025] Furthermore, the characteristic quantity characterizing the net load level of the system is composed of the difference between the load forecast value and the renewable energy output forecast value, and further includes its statistical distribution characteristics. The feature quantity characterizing the bidding behavior of the preset key units is used to filter the set of key units based on the frequency of occurrence of marginal units in historical clearing data, and the bidding information of the units in the set is extracted. The characteristic quantity that characterizes the overall supply and demand balance of the system is defined by the ratio of the difference between the total available generating capacity and the net load forecast to the net load forecast. The larger the ratio, the more relaxed the system is.

[0026] Furthermore, in the market state pattern segmentation step, the unsupervised learning algorithm used is the Gaussian mixture model clustering algorithm; the Gaussian mixture model can fit the complex probability distribution of features, thereby discovering naturally formed clusters in the data. Each price formation pattern obtained through clustering corresponds to a market operating state dominated by a specific type of marginal unit, giving the identified patterns a clear physical and economic explanation.

[0027] Furthermore, in the pattern probability prediction step, the constructed pattern probability prediction model is a multi-class probability prediction model. The input features of this model are predictive features for a specific future period. These predictive features are isomorphic to the multidimensional features used in the market state pattern segmentation step, ensuring the consistency of information transmission. The output of the model is the probability value of each price formation pattern belonging to a specific future time period. The probability value is a probability vector.

[0028] Furthermore, the multi-class probability prediction model is implemented using either a Softmax regression model or a gradient boosting decision tree model; both of these models can provide good classification probability calibration.

[0029] Furthermore, in the conditional price distribution prediction step, for any price formation pattern, the training data for its conditional probability prediction model is strictly derived from the subset of historical data that has been classified into that pattern. Conditional probability prediction models are implemented using quantile regression models or quantile regression forest models. These two models can directly output the conditional quantiles of variables, thus conveniently constructing a complete conditional probability distribution.

[0030] Furthermore, for any price formation mode, the input features of its conditional probability prediction model are a simplified feature set that is strongly correlated with the clearing price under that mode. The simplified feature set includes at least the net load level and the bid price of key units. Thus, overfitting is avoided through feature simplification, and the core influencing factors under that mode are highlighted.

[0031] Furthermore, the weighted synthesis method in the probability distribution synthesis step is as follows: The conditional probability distribution corresponding to each price formation pattern is multiplied by the probability of occurrence of that pattern obtained through the pattern probability prediction model. The weighted probability distributions of all patterns are then superimposed to form the final probability distribution.

[0032] Furthermore, the method is specifically applied to probabilistic price forecasting in a day-ahead clearing scenario of the electricity market containing a high proportion of intermittent renewable energy; in this scenario, the output of renewable energy is highly uncertain and marginal unit switching is frequent.

[0033] Through the above steps, the pattern partitioning step decomposes the complex nonlinear global problem; the pattern probability prediction step enables forward-looking judgment of future market conditions; the conditional distribution prediction step achieves high-precision learning within relatively simple linear subspaces; and the final synthesis step fully restores the overall uncertainty of the market. This not only improves prediction accuracy, especially the ability to capture price tail risks, but also endows the model with interpretability through the pattern concept, enabling decision-makers to clearly understand the driving factors behind price fluctuations and make reasonable decisions. It achieves more accurate and reliable probabilistic predictions of electricity market clearing prices, especially in environments with a high proportion of renewable energy. This invention uses the structural state of market operation as the hub of modeling, and through a discretization-conditionalization-resynthesis path, solves the problem that traditional models struggle to reflect the multi-peak characteristics of price probability distributions.

[0034] Example 2 Optionally, the present invention provides another embodiment, which provides a probabilistic prediction method for electricity market clearing prices, and the specific implementation process is as follows: Step S101: Market Status Pattern Classification First, a historical market operation database will be constructed, collecting electricity market operation data for at least the past three years, including: Hourly net load data: Load forecast minus the forecast output of renewable energy sources such as wind and solar power; Key unit pricing data: Select pricing curves for 5-8 key units with a historical marginal probability exceeding 20%; System supply-demand ratio data: the ratio of available generating capacity to net load; After standardizing the above multidimensional features, a Gaussian Mixture Model (GMM) was used for cluster analysis; the probability density function of the GMM is expressed as: p(x|θ) = Σ_{k=1}^K π_k N(x|μ_k, Σ_k) In this formula: x represents the standardized feature vector; K is the number of clusters, and the optimal value of K is determined by the Bayesian information criterion (usually 4-6). π_k is the mixing coefficient of the k-th Gaussian distribution, satisfying Σπ_k=1; μ_k is the mean vector of the k-th Gaussian distribution; Σ_k is the covariance matrix of the k-th Gaussian distribution; N(x|μ_k, Σ_k) represents a multivariate Gaussian distribution; By iteratively solving the model parameters using the expectation-maximization algorithm, historical market conditions are divided into typical patterns such as "baseload dominance", "mid-load competition", "peak load shortage", and "renewable energy surplus".

[0035] Step S102: Pattern Probability Prediction Construct a multi-class probability prediction model based on softmax regression. The model input features include: Forecast the net load curve at 96 points (15-minute intervals) on the day of the forecast and its statistical characteristics (mean, variance, peak-to-valley difference); Predict the price curve for key generating units submitted on the predicted day; Predict daily system reserve rate, tie-line power, and other supply and demand characteristics; The Softmax regression model is represented as: P(y=i|x) = exp(w_i^T x + b_i) / Σ_{j=1}^K exp(w_j^T x + b_j) In this formula: x is the input feature vector; y represents the pattern category, i∈{1,2,...,K}; w_i is the weight vector of the i-th class; b_i is the bias term of the i-th class; P(y=i|x) is the probability that input x belongs to the i-th class; The model training uses the cross-entropy loss function and optimizes the parameters through gradient descent to output the probability distribution of different price formation patterns for each time period of the predicted day.

[0036] Step S103: Predicting the distribution of conditional prices For each price formation pattern, all samples belonging to that pattern are selected from historical data, and conditional probability prediction models are trained for each pattern. Taking the "peak load shortage" model as an example, quantile regression forest is used to predict the conditional distribution. For a given set of conditional variables X and quantile levels τ∈(0,1), the conditional quantile Q_τ(Y|X) is estimated through the following steps: Construct B regression trees, and each tree draws samples from the training set using a bootstrap sampling method; For each tree, the loss function is minimized by recursively selecting features and split points through binary search; For a test sample x, record the output values ​​of all training samples in the leaf nodes of each tree; Conditional quantile estimates are empirical quantiles of the sample output values ​​across all trees; The model input features are simplified to: Net load forecast and its first difference; The pricing level of marginal units; System spinning reserve rate; The output is the 10%, 20%, ..., 90% quantiles of the clearing price under this mode, thus constructing the complete conditional probability distribution.

[0037] Step S104: Probability Distribution Synthesis Based on the law of total probability, the conditional distributions of each pattern are weighted and synthesized; for the t-th time period of the prediction day, the final price probability density function is: f_t(p) = Σ_{i=1}^KP(M_i|X_t) · f_{t,i}(p|M_i) In this formula: f_t(p) is the probability density function of the final price in time period t; P(M_i|X_t) is the probability of belonging to pattern M_i in time period t, which comes from step S102; f_{t,i}(p|M_i) is the price probability density function under pattern M_i, derived from step S103; K represents the total number of patterns; By using kernel density estimation, the quantiles are transformed into a continuous probability density function, ultimately yielding the probability distribution of the clearing price at 96 points on the prediction day.

[0038] Example 3 Optionally, the present invention provides another embodiment. Based on embodiment 2, in a power market with a high proportion of wind and solar power, this embodiment optimizes the above method as follows: Net load uncertainty modeling: Using a scenario-based analysis method, N net load scenarios are generated, each with a probability of π_n; the pattern probability prediction is then corrected as follows: P(M_i|X) = Σ_{n=1}^N π_n · P(M_i|X_n) Where X_n represents the feature vector under the nth net load scenario.

[0039] Temporal correlation processing for conditional distribution forecasts: The Copula function is used to characterize the dependency structure between prices in different time periods; for adjacent time periods t and t+1, the joint distribution function is expressed as: F(p_t, p_{t+1}|M_i) = C(F_t(p_t|M_i), F_{t+1}(p_{t+1}|M_i)) Where C is a Copula function, and either t-Copula or ClaytonCopula is selected to suit the peak-and-fat tail characteristics of electricity prices.

[0040] Extreme price handling mechanism: When the probability of a certain pattern is predicted to exceed a threshold (such as 30%) and the pattern corresponds to extreme high prices or negative electricity prices, Monte Carlo simulation is initiated to make a refined prediction and obtain more accurate heavy-tailed distribution characteristics through multiple sampling.

[0041] Example 4 Optionally, the present invention provides another embodiment, in which the specific system implementation of the present invention includes the following components: Data preprocessing module: The 3σ criterion was used to remove outliers from historical data; Missing data is handled using multiple interpolation methods; Feature engineering includes lagged variable construction, moving averages, etc.

[0042] The model training and validation process is as follows: Time series cross-validation was used, with the training and test sets divided in a 7:3 ratio. Hyperparameter optimization employs a Bayesian optimization method to search for key parameters: GMM clustering number K: range [3, 8]; Quantile regression of the number of trees in the forest: range [100, 500]; Softmax regression regularization coefficient: range [0.01, 1]; Real-time prediction process: Get market data at a fixed time every day for the forecast date (96 points). Perform pattern probability prediction and output the pattern distribution for each time period; Conditional distribution predictions are performed in parallel under each mode; Synthesize the final probability distribution and generate risk indicators; The forecast results are pushed to the decision-making systems of market participants.

[0043] Through the above specific implementation methods, the present invention can effectively capture the multi-peak distribution characteristics of electricity market clearing prices, providing market participants with more accurate risk assessment and decision support.

[0044] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A probabilistic method for forecasting market clearing prices in an electricity market, characterized by: The method comprises the following steps: a market state mode division step: based on historical market operation data, multi-dimensional features related to the clearing price formation mechanism are extracted, and an unsupervised learning algorithm is used to divide the historical market state into a limited number of discrete price formation modes; a mode probability prediction step: based on the results of the market state mode division step, a mode probability prediction model is constructed and trained to output the probability of each price formation mode in a future specific period; a conditional price distribution prediction step: for each price formation mode, a conditional probability prediction model is constructed and trained to output the conditional probability distribution of the market clearing price under the condition that the mode occurs; a probability distribution synthesis step: for a future specific period to be predicted, the mode probability output by the mode probability prediction step and the corresponding conditional probability distribution output by the conditional price distribution prediction step are synthesized by a weighted synthesis method to calculate the final probability distribution of the market clearing price in the period.

2. The probabilistic method for forecasting market clearing price according to claim 1, wherein: In the market state mode division step, the multi-dimensional features related to the clearing price formation mechanism at least include: a feature quantity representing the system net load level; a feature quantity representing the bidding behavior of a preset key unit; a feature quantity representing the overall supply-demand slackness of the system.

3. A probabilistic method for forecasting market clearing prices according to claim 2, characterized in that: The feature quantity representing the system net load level is composed of the difference between the load forecast value and the renewable energy output forecast value, and further includes the statistical distribution characteristics thereof; The feature quantity representing the bidding behavior of the preset key unit is obtained by screening the key unit set based on the occurrence frequency of marginal units in historical clearing data, and extracting the bidding information of the units in the set; The feature quantity representing the overall supply-demand slackness of the system is defined by the ratio of the difference between the total available generation capacity of the system and the net load forecast value to the net load forecast value.

4. The probabilistic method for forecasting market clearing price according to claim 1, wherein: In the market state mode division step, the unsupervised learning algorithm used is a Gaussian mixture model clustering algorithm; Each price formation mode obtained by clustering corresponds to a market operation state dominated by a specific marginal unit type.

5. The probabilistic method for forecasting market clearing price according to claim 1, wherein: In the mode probability prediction step, the constructed mode probability prediction model is a multi-class probability prediction model; The input features of the model are the predictive features of a future specific period, which are the same as the multi-dimensional features used in the market state mode division step; The output of the model is the probability value of the future specific period belonging to each price formation mode.

6. A probabilistic method for forecasting market clearing prices according to claim 5, characterized in that: The multi-class probability prediction model is implemented using a Softmax regression model or a gradient boosting decision tree model.

7. The probabilistic method for forecasting market clearing price of electricity as claimed in claim 1 wherein: In the conditional price distribution prediction step, the training data of the conditional probability prediction model for any price formation mode strictly comes from the data subset divided into the mode in the historical data; The conditional probability prediction model is implemented using a quantile regression model or a quantile regression forest model.

8. A probabilistic method for forecasting market clearing prices according to claim 7, characterized in that: For any price formation mode, the input features of the conditional probability prediction model are the simplified feature set strongly related to the clearing price under the mode, which at least includes the net load level and the key unit bidding.

9. The probabilistic method for forecasting market clearing price of electricity according to claim 1, wherein: The weighted synthesis method in the probability distribution synthesis step is as follows: The condition probability distribution corresponding to each price formation pattern is multiplied by the occurrence probability of the pattern via the pattern probability prediction model, and all the weighted probability distributions under all patterns are superimposed to form the final probability distribution.

10. A probabilistic method for forecasting market clearing prices of an electricity market according to any of claims 1-9, characterized in that: The method is particularly applied to the probabilistic price prediction of the day-ahead market clearing scenario of the power market containing a high proportion of intermittent renewable energy.

Images