CN-121981756-A - Day-ahead electricity price prediction method and device based on real-time market demand prediction

CN121981756ACN 121981756 ACN121981756 ACN 121981756ACN-121981756-A

Abstract

A day-ahead electricity price prediction method and device based on real-time market demand prediction. The method comprises the steps of obtaining market game expected indexes and traditional market data in a historical trading period, constructing a fusion feature matrix based on the market game expected indexes and the traditional market data in the historical trading period, matching the fusion feature matrix with preset historical day-ahead electricity price data to generate a training data set, training an integrated probability prediction model according to the training data set to obtain a target integrated probability prediction model, obtaining the market game expected indexes and the traditional market data in the period to be predicted to construct a feature matrix to be predicted, inputting the feature matrix to be predicted into the target integrated probability prediction model to obtain a prediction processing result, and generating day-ahead electricity price single-point predicted values and day-ahead electricity price predicted intervals according to the prediction processing result. By implementing the scheme, more comprehensive and reliable price signals and risk quantization information can be provided for operation optimization decisions of the power system.

Inventors

PAN YINGCHAO
DUAN XIAOHAN

Assignees

北京如实智慧电力科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (10)

1. A day-ahead electricity price prediction method based on real-time market demand prediction, characterized by being applied to a server, the method comprising: acquiring market game expected indexes and traditional market data in a historical trading period; constructing a fusion feature matrix based on the market gaming expected indicators and the traditional market data in the historical trading period; matching the fusion feature matrix with corresponding preset historical day-ahead electricity price data in the historical transaction period to generate a training data set; Training a preset integrated probability prediction model according to the training data set to obtain a target integrated probability prediction model; obtaining market game expected indexes and traditional market data of a period to be predicted so as to construct a feature matrix to be predicted; inputting the feature matrix to be predicted into the target integrated probability prediction model to obtain a prediction processing result; And generating a single-point predicted value of the day-ahead electricity price corresponding to the period to be predicted and a day-ahead electricity price prediction interval based on a preset confidence level according to the prediction processing result.
2. The method of claim 1, wherein the constructing a fused feature matrix based on the market gaming expectation indicators and the legacy market data over the historical trading period comprises: extracting dynamic evolution features used for representing time dynamic changes from the market game expected indexes in the historical transaction period; Extracting contextual statistical features characterizing statistical distribution characteristics from the traditional market data over the historical trading period; Respectively inputting the dynamic evolution features and the context statistical features into a preset attention computing network to generate dynamic feature attention weight vectors and context feature attention weight vectors; Multiplying the dynamic evolution feature by the dynamic feature attention weight vector to obtain a recalibrated dynamic feature, and multiplying the context statistical feature by the context feature attention weight vector to obtain a recalibrated context feature; and splicing the recalibration dynamic features and the recalibration context features in a preset feature dimension to generate the fusion feature matrix.
3. The method according to claim 1, wherein training the preset integrated probability prediction model according to the training data set to obtain the target integrated probability prediction model comprises: Dividing the training data set into a training set and a testing set according to a preset proportion; according to the training set, training to obtain a plurality of base prediction models belonging to different preset model types; According to the test set, independent prediction performance evaluation is carried out on a target base prediction model, a prediction error index corresponding to the target base prediction model is generated, and the target base prediction model is any base prediction model; Calculating the integration weight of the target base prediction model according to the prediction error index, wherein the integration weight is in inverse relation with the prediction error index, and performing normalization processing on all the integration weights; And combining all the base prediction models with the normalized integration weights to construct the target integration probability prediction model.
4. The method of claim 3, wherein the performing independent prediction performance evaluation on the target base prediction model according to the test set, generating a prediction error indicator corresponding to the target base prediction model, includes: inputting the fusion feature matrix in the test set to the target base prediction model to obtain a model prediction electricity price corresponding to each data point in the test set; Calculating an absolute value of a prediction bias based on the model predicted electricity price and the preset historical day-ahead electricity price data corresponding to the data points for each data point in the test set; And carrying out arithmetic average on absolute values of the prediction deviations corresponding to all the data points in the test set, and taking the obtained average value as the prediction error index corresponding to the base prediction model.
5. The method according to claim 1, wherein the generating, according to the prediction processing result, a single-point predicted value of a day-ahead power price and a day-ahead power price prediction interval based on a preset confidence level, which correspond to the period to be predicted, includes: Analyzing the prediction processing result into a plurality of quantile predicted values corresponding to a plurality of preset quantile levels, wherein one preset quantile level corresponds to one quantile predicted value; Selecting a quantile predicted value corresponding to a preset reference quantile in the quantile predicted values as the day-ahead electricity price single-point predicted value; Determining a lower quantile level and an upper quantile level according to the preset confidence level; and respectively selecting quantile predicted values corresponding to the lower quantile level and the upper quantile level in the quantile predicted values as the lower limit and the upper limit of the day-ahead electricity price predicted interval.
6. The method of claim 5, wherein determining the lower quantile level and the upper quantile level based on the preset confidence level comprises: performing halving operation on the preset confidence level to obtain a confidence interval half-width; subtracting the confidence interval half width from the preset reference quantile to obtain a difference value, and taking the difference value as the lower quantile level; And adding the preset reference quantile to the confidence interval half width to obtain a sum value, and taking the sum value as the upper quantile level.
7. A method according to claim 3, characterized in that the method further comprises: acquiring real day-ahead electricity price data corresponding to the period to be predicted, and calculating a real-time prediction error between the day-ahead electricity price single-point predicted value and the real day-ahead electricity price data; Calculating statistical distribution distances between the feature matrix to be predicted and all the fusion feature matrices in the training data set to generate a feature outlier index; When the real-time prediction error exceeds a preset error threshold and the characteristic outlier index is lower than a preset outlier threshold, judging a new data sample formed by the to-be-predicted characteristic matrix and the real day-ahead electricity price data as an effective incremental sample, and adding the effective incremental sample into the training data set to form an updated training data set; And carrying out updating training on the target integrated probability prediction model based on the updated training data set so as to obtain the updated target integrated probability prediction model.
8. An electronic device comprising a processor, a memory, a user interface, and a network interface, the memory for storing instructions, the user interface and the network interface each for communicating with other devices, the processor for executing instructions stored in the memory to cause the electronic device to perform the method of any of claims 1-7.
9. A computer readable storage medium storing instructions which, when executed, perform the method of any one of claims 1-7.
10. A computer program product, characterized in that the computer program product, when run on an electronic device, causes the electronic device to perform the method of any of claims 1-7.

Description

Day-ahead electricity price prediction method and device based on real-time market demand prediction Technical Field The application relates to the technical field of electric power market prediction and risk management, in particular to a day-ahead electricity price prediction method and device based on real-time market demand prediction. Background In the prior art, a model driven by time series data is mainly adopted for the day-ahead electricity price prediction. Specifically, a time sequence prediction model is generally trained by acquiring and utilizing macro market data such as historical electricity prices, system loads, meteorological parameters and the like, and future electricity prices are predicted by mining statistical rules in the data. However, the above-described method has significant drawbacks. This is because the existing methods rely entirely on statistical laws of historical data to extrapolate, failing to model strong price conductance relationships and expected gaming dynamics between the day-ahead market and the real-time market based on the physical operating constraints of the power system. The failure of the critical game expectation index, which is a real-time market signal (the signal is essentially the marginal value of the instantaneous physical unbalance state of the power system), leads to serious failure of the prediction model based on pure history statistics when the system state suddenly changes (such as the severe fluctuation of the output of renewable energy sources, the failure of a big unit and the like). As a result, the predicted electricity price cannot accurately reflect the future real operation situation of the system, and thus, reliable price signal input cannot be provided for critical prior operation optimization decisions of the power system, such as unit combination, safety constraint economic dispatch, standby capacity dispatch and the like. Disclosure of Invention In order to solve the technical problems, the application provides a day-ahead electricity price prediction method and device based on real-time market demand prediction. The application provides a day-ahead electricity price prediction method based on real-time market demand prediction, which adopts the following technical scheme: acquiring market game expected indexes and traditional market data in a historical trading period; constructing a fusion feature matrix based on the market gaming expected indicators and the traditional market data in the historical trading period; matching the fusion feature matrix with corresponding preset historical day-ahead electricity price data in the historical transaction period to generate a training data set; Training a preset integrated probability prediction model according to the training data set to obtain a target integrated probability prediction model; obtaining market game expected indexes and traditional market data of a period to be predicted so as to construct a feature matrix to be predicted; inputting the feature matrix to be predicted into the target integrated probability prediction model to obtain a prediction processing result; And generating a single-point predicted value of the day-ahead electricity price corresponding to the period to be predicted and a day-ahead electricity price prediction interval based on a preset confidence level according to the prediction processing result. By adopting the technical scheme, market indexes reflecting the real-time physical state and game expectation of the system are introduced, the market indexes are fused with traditional market data to construct a feature matrix, and training and prediction are performed based on an integrated probability model. The method effectively models the price conduction mechanism and game dynamics between the day-ahead market and the real-time market, and remarkably improves the prediction robustness and accuracy under the system state mutation scene such as renewable energy output fluctuation. The output probabilistic prediction result (single point value and confidence interval) can provide more comprehensive and reliable price signal and risk quantization information for the operation optimization decision of the power system. Optionally, the constructing a fusion feature matrix based on the market gaming expected indicators and the traditional market data in the historical trading period includes: extracting dynamic evolution features used for representing time dynamic changes from the market game expected indexes in the historical transaction period; Extracting contextual statistical features characterizing statistical distribution characteristics from the traditional market data over the historical trading period; Respectively inputting the dynamic evolution features and the context statistical features into a preset attention computing network to generate dynamic feature attention weight vectors and context feature attention weight vectors; Multiplying the dyn