Day-ahead-real-time electricity market price difference prediction method considering global sensitivity analysis
Through global sensitivity analysis and CatBoost classifier, combined with Sobol effect index, we predict the recent and real-time price difference of the power market, which solves the market uncertainty and trading strategy formulation problems faced by power sales companies, improves the prediction accuracy and robustness, and provides More comprehensive market strategy support.
Patent Information
- Application Number
- CN202410956552.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-15
AI Technical Summary
Power sales companies face risks brought by market price uncertainty and user electricity consumption volatility in the power market. Existing research is mostly focused on a single market (someday market or real-time market), and there is a lack of effective price spread prediction methods to reduce the price difference. Determine and develop trading strategies.
A method that considers global sensitivity analysis is adopted, and the gradient-enhancing decision tree library classifier (CatBoost) is generated through training, combined with Sobol effect index calculation, identify and sort the impact of each feature on the output variable, and output prediction results, including grid operation data, Temperature, coal price and time labels.
It improves prediction accuracy and model robustness, reduces the impact of noise, provides more comprehensive prediction information and uncertainty analysis, and helps power sales companies formulate more reasonable market strategies.
Smart Images

Figure CN118967241A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a prediction technology applied in a power market transaction process, and specifically to a prediction technology for a price difference between a day-ahead market and a real-time market in a power market. Background Art
[0002] The continuous deepening of the power system reform has provided good conditions for the establishment of a competitive market on the power sales side. In the new retail market, power sales companies purchase electricity from the power generation side through market transactions and sell electricity to users on behalf of the power grid companies, which has improved the economic benefits of the user side to a certain extent.
[0003] Power sales companies purchase electricity from different markets through centralized bidding or bilateral negotiation, and then sell it to users at a suitable price based on user demand, earning profits from the difference between purchase and sale. Since both power sales companies and users have the right to choose to purchase electricity, power sales companies have to face both the uncertainty of market prices and the volatility of user electricity consumption, which brings great risks to power sales companies in power trading.
[0004] In the short term, the electricity market can be divided into the day-ahead market and the real-time market according to the time when the transaction occurs. Most of the transactions of power sales companies also occur in these two markets. Predicting the positive or negative difference between the day-ahead market price and the real-time market price can reduce the uncertainty faced by power sales companies to a certain extent, and also help power sales companies formulate trading strategies. However, most of the existing studies focus on one of the day-ahead market or the real-time market, and the few studies that consider the positive or negative price difference also use more traditional mathematical models. With the development of machine learning methods, the use of machine learning models for price difference prediction will bring higher accuracy. At the same time, the use of global sensitivity analysis methods can also better compare the weights of each feature, which is conducive to power sales companies to better respond to market changes. Summary of the invention
[0005] The purpose of this invention is to improve the ability of power sales companies to deal with uncertainty and effectively formulate trading strategies in the power market. A method for predicting the price difference between the day-ahead price and the real-time price in the power market and giving a ranking of the impact of characteristics is proposed.
[0006] The technical solution of the present invention is as follows:
[0007] A day-ahead real-time electricity market price difference prediction method considering global sensitivity analysis is characterized by: S1, inputting training data of actual power grid operation to train and generate a gradient boosting decision tree library classifier; S2, using a global sensitivity analysis method to calculate the ranking and weight of the impact of each feature on the output variable; S3, outputting the prediction result based on the power grid operation data, temperature, coal price and time label of the period to be predicted.
[0008] The specific steps are as follows:
[0009] (1) Input the actual operation training data of the power grid and perform minimum and maximum normalization processing on each dimension of the data, mapping the original data to the interval [0,1], including the actual operation data of the power grid, temperature, coal price, corresponding time label, and the positive and negative data of the price difference between the day-ahead market and the real-time market;
[0010] (2) Generate a CatBoost classifier based on input data training; CatBoost is an open source gradient boosting decision tree library;
[0011] (3) Generate a sampling matrix B for calculating the Sobol index;
[0012] First, use the Sobol sequence to generate a basic matrix A of size N*2D, where N is the number of samples and D is the feature dimension;
[0013] Each row of matrix A is recorded as
[0014] A i =[a i,1 ,a i,2 ,…,a i,D ,b i,1 ,b i,2 ,…,b i,D ],i=1,2,…N (1)
[0015] where a i,1 ,a i,2 ,…,a i,D represents the first D elements of the i-th row in matrix A, b i,1 ,b i,2 ,…,b i,D Represents the last D elements of the i-th row in matrix A. The data from the 1st to the Nth rows of matrix A are the data from the N+1th to the 2Nth rows arranged in a Sobol sequence with 2D data per row.
[0016] Afterwards, for each row of data A i , the sampling matrix B is obtained by cross-pollination as follows i .
[0017]
[0018] Here a1b1 represents the data in the basic matrix. The division into a1b1 is for the sake of clarity in the subsequent description of the sampling matrix generation process. It is not necessarily related to the aforementioned training data. The purpose of generating this matrix is to obtain matrix C, that is, input the sampling matrix B as similar input data into the classifier, and then obtain matrix C. The Sobol effect index of each order is calculated according to matrix C. The 1st to Nth row data of matrix A are the N+1th to 2Nth row data arranged in Sobol sequence with 2D data per row.
[0019] B i The first and last rows of A i The first and second half of the 2nd to (D+1)th row will be A i The corresponding elements in the second half of are replaced with the corresponding positions in the first half, that is, for the sampling matrix B i The jth row of B ij ,have
[0020] B ij =[a i,1 ,a i,2 ,…,a i,j-2 ,b i,j-1 ,a i,j ,…a i,D ],j=2,3,…,D+1 (3)
[0021] Rows (D+2) to (2D+1) will be A i The corresponding elements in the first half are replaced with the corresponding positions in the second half, that is, for the sampling matrix B i The jth row of B ij ,have
[0022] B ij =[b i,1 ,b i,2 ,…,b i,j-2 ,a i,j-1 ,b i,j ,…b i,D ],j=D+2,D+3,…,2D+1 (4)
[0023] Then we get a sampling matrix B of size (2D+2)N*D:
[0024] B=[B1,B2,B3,…,B N ] T (5)
[0025] (4) Input matrix B into the Catboost classifier to obtain the output matrix C;
[0026] (5) Calculate the Sobol first-order effect index, second-order effect index and total effect index of each feature based on matrix C; the calculation formula of Sobol first-order effect index is:
[0027]
[0028] Among them, S i Represents the i-th input variable X i The first-order effect index indicates the direct impact of a single input variable on the output variable; E(Y|X i ) is the expected value of the output variable given the i-th input variable; Y is the output variable; Var(·) represents the variance calculation.
[0029] The calculation formula of Sobol second-order effect index is:
[0030]
[0031] Among them, S ij represents the second-order effect index between the i-th and j-th output variables, indicating the impact of the interaction of two input variables on the output variable; E(Y|X i ,X j ) is the expected value of the output variable under the condition of the i-th and j-th input variables;
[0032] The calculation formula of Sobol total effect index is:
[0033]
[0034] in, is the total effect index of the i-th input variable, indicating the total impact of a single input variable on the output variable; X ~i are all other input variables except the i-th input variable; E(Y|X ~i ) is the expected value of the output variable given all input variables except the i-th input variable;
[0035] (6) Sort the calculated Sobol first-order effect index, second-order effect index and total effect index, and output the ranking of various types of influence of the input variable on the output variable and the degree of influence. The ranking is in descending order. The larger the effect index, the greater the influence of the input variable on the output variable.
[0036] (7) The grid operation data, temperature, coal price and time label of the predicted date are input into the trained CatBoost classifier, and the classification prediction result is output. The output result may be -1 or 1. When the result is 1, it means that the day-ahead electricity price is greater than the real-time electricity price. When the result is -1, it means that the day-ahead electricity price is less than the real-time electricity price.
[0037] The method for predicting the day-ahead and real-time power market price spread considering global sensitivity analysis is characterized by:
[0038] The actual operation data of the power grid include the centralized load, Class A power output, Class B power output, local power output, external support power and interconnection line power; the coal price is the coal guidance price; the time label includes the week, month, day, hour and minute; Class A power refers to the power that has not participated in market transactions, and Class B power refers to the power that has participated in market transactions.
[0039] The method for predicting the day-ahead and real-time power market price spread considering global sensitivity analysis is characterized by:
[0040] The training generates a CatBoost classifier by first training a base learner from the initial training set, and then adjusting the distribution of training samples according to the performance of the base learner so that the training samples that the previous base learner got wrong have greater weights in the subsequent period, and the weight range is [0,1]. Then, the next base learner is trained based on the adjusted sample distribution, and the process is repeated until the base learner data reaches a predetermined value T, and finally the T base learners are weightedly combined to obtain the final strong learner.
[0041] The larger the T value is, the more complex the entire classifier is and the longer the training time is. In practice, the T value is usually determined by comprehensively considering the complexity of the problem and the available computing resources.
[0042] The method for predicting the day-ahead and real-time power market price spread considering global sensitivity analysis is characterized by:
[0043] The sampling matrix B, the 1st to 13th input variables are: centralized load, Class A power output, Class B power output, local power output, external support power, interconnection line power, temperature, coal price, week sequence, month, day, hour, minute, and the output variable is the positive or negative price difference between the day-ahead market and the real-time market.
[0044] Advantages of the present invention:
[0045] 1. Improve prediction accuracy:
[0046] Through global sensitivity analysis, this method can identify and select the input features that have the greatest impact on the prediction results, thereby reducing the impact of noise and irrelevant data on the prediction accuracy. Ensemble learning can more comprehensively capture the complex patterns and nonlinear relationships in the data by combining the prediction results of multiple base learners, further improving the accuracy of the prediction.
[0047] 2. Enhance model robustness:
[0048] Multiple base learners in ensemble learning can complement each other, reduce the risk of overfitting of a single model, and improve the generalization ability of the model. When some base learners are disturbed by abnormal data or noise, other base learners can still provide stable prediction results, thereby enhancing the robustness of the overall model.
[0049] 3. Strong adaptability:
[0050] This method does not rely on a specific electricity market structure or data distribution, and can be flexibly applied to electricity market price spread forecasting in different regions. By adjusting the type and number of base learners, this method can adapt to data sets of different complexity and size to achieve efficient forecasting.
[0051] 4. Provide comprehensive forecast information:
[0052] The ensemble learning method can output the prediction results and confidence levels of multiple base learners, providing decision makers with more comprehensive prediction information and uncertainty analysis. This helps decision makers better understand the reliability and potential risks of prediction results and formulate more reasonable market strategies.
[0053] 5. Easy to implement and expand:
[0054] The method is based on existing machine learning and statistical analysis techniques and is easy to implement and deploy.
[0055] As new forecasting algorithms and data sources emerge, the method can be easily expanded and updated to keep the forecasts advanced and accurate.
[0056] The present invention firstly trains a gradient boosting decision tree library classifier using existing data, then uses a global sensitivity analysis method to obtain the ranking and weight of the impact of each feature on the output variable, and finally outputs the prediction result based on the grid operation data, temperature, coal price and time label of the forecast period, which is conducive to power sales companies to better cope with uncertainties in the market and effectively formulate trading strategies.
[0057] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 It is a flow chart of a method for predicting day-ahead and real-time electricity market price spreads taking into account global sensitivity analysis according to an embodiment of the present invention.
[0059] Figure 2 The embodiment of the present invention is a schematic diagram of the Boost method.
[0060] Figure 3 It is a graph of the total effect index of each feature obtained by using the present invention. DETAILED DESCRIPTION
[0061] The present invention is further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.
[0062] It should be noted that the directional terms such as left, right, up, down, top, and bottom in this embodiment are merely relative concepts or are based on the normal use status of the product, and should not be considered as restrictive.
[0063] A method for predicting the price difference between day-ahead and real-time electricity markets considering global sensitivity analysis.
[0064] The specific steps are as follows:
[0065] (1) Input training data and normalize each dimension of data, including actual grid operation data, temperature, coal price, corresponding time label, and positive and negative price difference data between the day-ahead market and the real-time market. The actual grid operation data includes unified load, Class A power output, Class B power output, local power output, external support power, and interconnection line power; coal price is the coal guidance price; time label includes weekday, month, day, hour, and minute. Class A power refers to power that has not participated in market-based transactions, and Class B power refers to power that has participated in market-based transactions.
[0066] (2) Generate a CatBoost classifier based on input data training. CatBoost is an open source gradient boosting decision tree library that specifically optimizes the ability to process category features compared to traditional gradient boosting decision trees. It is also an implementation of the Boost method in ensemble learning. The schematic diagram of the Boost method is shown in the figure below. Figure 2 As shown. The whole method first trains a base learner from the initial training set, and then adjusts the distribution of training samples according to the performance of the base learner, so that the training samples that the previous base learner did wrong have a larger weight in the subsequent period, and the weight range is [0,1]. Then, the next base learner is trained based on the adjusted sample distribution, and repeated until the base learner data reaches the pre-specified value T, and finally the T base learners are weighted and combined to obtain the final strong learner. The larger the T value, the more complex the entire classifier, and the longer the training time required. In practice, the T value is usually determined by comprehensively considering the complexity of the problem and the available computing resources. The weights of the entire CatBoost training process are implicitly expressed, and then the output results are weighted and summed, so there is no specific numerical range. The larger the sample weight, the more reflected in the learning results of the next base learner. From the perspective of the sample set, a certain sample is given a larger weight, and the weight range is also in the range of [0,1] (because the sum of the sample weights of the entire sample set is 1).
[0067] (3) Generate a sampling matrix B for calculating the Sobol index. In the present invention, the first to thirteenth input variables are: centralized load, Class A power output, Class B power output, local power output, external support power, interconnection line power, temperature, coal price, week, month, day, hour, minute, and the output variable is the positive or negative price difference between the day-ahead market and the real-time market. First, use the Sobol sequence to generate a basic matrix A of size N*2D, where N is the number of sampling samples (taken as 2048 in the present invention) and D is the feature dimension (D is taken as 13 in the present invention). Let each row of matrix A be
[0068] A i =[a i,1 ,a i,2 ,…,a i,D ,b i,1 ,b i,2 ,…,b i,D ],i=1,2,…N (1)
[0069] where a i,1 ,a i,2 ,…,a i,D represents the first D elements of the i-th row in matrix A, b i,1 ,b i,2 ,…,b i,D Represents the last D elements of the i-th row in matrix A. The data from the 1st to the Nth rows of matrix A are the data from the N+1th to the 2Nth rows arranged in a Sobol sequence with 2D data per row.
[0070] Afterwards, for each row of data A i , the sampling matrix B is obtained by cross-pollination as follows i .
[0071]
[0072] The number of sampling samples N is selected as 2048 and the feature dimension D is selected as 13.
[0073] Here, a1b1 represents the data in the basic matrix. The division into a1b1 is for the purpose of making the subsequent description of the sampling matrix generation process clearer and has no necessary connection with the aforementioned training data. The purpose of generating this matrix is to obtain the matrix C, that is, the sampling matrix B is input into the classifier as similar input data, and then the matrix C is obtained, and the Sobol effect index of each order is calculated according to the matrix C. The data from the 1st row to the Nth row of the matrix A are the data from the N+1th to the 2Nth row arranged by the Sobol sequence with 2D data per row (that is, removing the first N rows, which are the 2049th to 4096th rows of data in the present invention)
[0074] B i The first and last rows of A iThe first and second half of the 2nd to (D+1)th row will be A i The corresponding elements in the second half of are replaced with the corresponding positions in the first half, that is, for the sampling matrix B i The jth row of B ij ,have
[0075] B ij =[a i,1 ,a i,2 ,…,a i,j-2 ,b i,j-1 ,a i,j ,…α i,D ],j=2,3,…,D+1 (3)
[0076] Rows (D+2) to (2D+1) will be A i The corresponding elements in the first half are replaced with the corresponding positions in the second half, that is, for the sampling matrix B i The jth row of B ij ,have
[0077] B ij =[b i,1 ,b i,2 ,…,b i,j-2 ,a i,j-1 ,b i,j ,…b i,D ],j=D+2,D+3,…,2D+1 (4)
[0078] Then we get a sampling matrix B of size (2D+2)N*D:
[0079] B=[B1,B2,B3,…,B N ] T (5)
[0080] (4) Input matrix B into the Catboost classifier to obtain the output matrix C.
[0081] (5) Calculate the Sobol first-order effect index, second-order effect index and total effect index of each feature based on matrix C. The calculation formula of Sobol first-order effect index is:
[0082]
[0083] Among them, S i Represents the i-th input variable X i The first-order effect index indicates the direct impact of a single input variable on the output variable; E(Y|X i ) is the expected value of the output variable given the i-th input variable; Y is the output variable; Var(·) represents the variance calculation.
[0084] The calculation formula of Sobol second-order effect index is:
[0085]
[0086] Among them, S ij represents the second-order effect index between the i-th and j-th output variables, indicating the impact of the interaction of two input variables on the output variable; E(Y|X i ,X j ) is the expected value of the output variable conditional on the i-th and j-th input variables.
[0087] The calculation formula of Sobol total effect index is:
[0088]
[0089] in, is the total effect index of the i-th input variable, indicating the total impact of a single input variable on the output variable; X ~i are all other input variables except the i-th input variable; E(Y|X ~i ) is the expected value of the output variable given all input variables except the ith input variable.
[0090] (6) Sort the calculated Sobol first-order effect index, second-order effect index and total effect index, and output the ranking of various types of influence of the input variable on the output variable and the degree of influence. The ranking is in descending order. The larger the effect index, the greater the influence of the input variable on the output variable. The purpose of the ranking is to obtain the ranking of the influence of different input variables on the output variable. The overall output includes two parts, one is the prediction result, and the other is the influence of the input variable on the output variable. The seventh step outputs the first half, and the sixth step outputs the second half.
[0091] (7) Input the grid operation data, temperature, coal price and time tag of the date to be predicted, and output the prediction result.
[0092] The following are application examples:
[0093] Taking the actual electricity market transaction data of a certain city as an example for analysis, 366 consecutive days of data were selected, with one data point every 15 minutes. The training set and validation set were divided into training set and validation set in a ratio of 4:1. The effectiveness of the proposed prediction method was verified by comparing the accuracy of the validation set.
[0094] First, we input the actual operation data of the power grid, temperature, coal price, corresponding time label, and the positive and negative data of the price difference between the day-ahead market and the real-time market. The input matrix is shown in Table 1.
[0095] Table 1 Original input data
[0096]
[0097] Each column of data in the table represents the centralized load, Class A power output, Class B power output, local power output, external support power, interconnection line power, temperature, coal price, weekday, month, day, hour, minute and positive and negative price difference from left to right. The unit of centralized load, Class A power output, Class B power output, local power output, external support power and interconnection line power is megawatt, the unit of temperature is Celsius, and the unit of coal price is yuan / ton. In the positive and negative price difference data, -1 means that the market price on the day before is greater than the real-time market price, and 1 means that the real-time market price is greater than the market price on the day before.
[0098] After that, the CatBoost classifier is trained based on the data, and a sampling matrix is generated according to step (3). The sampling matrix is input into the previously trained CatBoost classifier to obtain an output matrix. In this process, the trained CatBoost classifier can be regarded as a black box that performs a certain transformation, and the input matrix is transformed to obtain the output matrix. The basic matrix generated by the Sobol sequence is shown in Table 2, with a size of 2048*26; the sampling matrix is shown in Table 3, with a size of 57344*13; and the output matrix is shown in Table 4, with a size of 57344*1. Since the output matrix is a column vector, for ease of presentation, Table 4 shows the transpose of the output matrix.
[0099] Table 2 Basic matrix data
[0100]
[0101] Table 3 Sampling matrix data
[0102]
[0103] Table 4 Output matrix data
[0104]
[0105] Afterwards, the Sobol first-order effect index, second-order effect index, and total effect index of each feature are calculated according to formulas (6), (7), and (8) and sorted in descending order to obtain the impact ranking of different input variables on the output variable, where the higher the ranking, the greater the impact of the input variable on the output variable. Finally, according to step (7), the power grid operation data, temperature, coal price, and time label of the date to be predicted are input into the CatBoost classifier to obtain the prediction result. The whole process finally outputs the effect index of each order and the corresponding ranking, as well as the prediction result of the date to be predicted.
[0106] The comparison method selects the currently more mature support vector machine (SVM) classifier, and compares the training time, prediction time and accuracy, as shown in Table 5. In addition, the method of the present invention can also obtain the degree of influence of different feature quantities on the output, such as Figure 2 shown.
[0107] Table 5 Comparison of the results of the present invention and SVM
[0108] method The present invention SVM Classifier Training time 70.92s 46.51s Prediction time 0.06s 18.72s Accuracy 86.30% 62.71%
[0109] It can be seen that although the training time of the present invention is slightly longer than that of the SVM classifier, it is significantly better than the SVM classifier in terms of prediction time and accuracy. At the same time, it can also give the degree of influence of different features on the output, thus illustrating the effectiveness of the present invention. Figure 3 The total effect index of different features on the output can also be seen in . The larger the total effect index, the greater the influence of the feature on the output. Figure 3 It can be seen that local power output, Class B power output and Class A power output are the three variables that have the greatest impact on the price difference, which is also a result that the SVM method cannot obtain.
[0110] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.
Claims
1. A method for predicting the price difference in the day-ahead and real-time electricity market considering global sensitivity analysis, characterized in that: S1. Input the training data of the actual operation of the power grid to train and generate the gradient boosting decision tree library classifier; S2. Use the global sensitivity analysis method to calculate the ranking and weight of the impact of each feature on the output variable; S3. Output the prediction result based on the power grid operation data, temperature, coal price and time label of the prediction period.
2. A method for predicting the day-ahead and real-time power market price spread considering global sensitivity analysis according to claim 1, characterized in that: Input the training data of the actual operation of the power grid, including the actual operation data of the power grid, temperature, coal price and time label, as well as the positive and negative data of the price difference between the day-ahead market and the real-time market.
3. A method for predicting the day-ahead and real-time power market price spread considering global sensitivity analysis according to claim 1 or 2, characterized in that: The training data of the actual operation of the input power grid is first normalized for each dimension of the data; the gradient boosting decision tree library classifier adopts the CatBoost classifier.
4. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 1 is characterized by: The Sobol global sensitivity analysis method is used to calculate the ranking and weight of the impact of each feature on the output variable. Finally, the prediction results are output based on the power grid operation data, temperature, coal price and time label of the forecast period, including the following steps: S2-1, generate a sampling matrix B for calculating the Sobol index; First, use the Sobol sequence to generate a basic matrix A of size N*2D, where N is the number of samples and D is the feature dimension; Each row of matrix A is recorded as A i =[a i,1 ,a i,2 ,…,a i,D ,b i,1 ,b i,2 ,…,b i,D ],i=1,2,...N where a i,1 ,a i,2 ,…,a i,D represents the first D elements of the i-th row in matrix A, b i,1 ,b i,2 ,…,b i,D Represents the last D elements of the i-th row in matrix A. The data from the 1st to the Nth rows of matrix A are the data from the N+1th to the 2Nth rows arranged in a Sobol sequence with 2D data per row. Afterwards, for each row of data A i , the sampling matrix B is obtained by cross-pollination as follows i . B i The first and last rows of A i The first and second half of the 2nd to (D+1)th row will be A i The corresponding elements in the second half of are replaced with the corresponding positions in the first half, that is, for the sampling matrix B i The jth row of B ij ,have B ij =[a i,1 ,a i,2 ,…,a i,j-2 ,b i,j-1 ,a i,j ,…a i,D ],j=2,3,…,D+1 Rows (D+2) to (2D+1) will be A i The corresponding elements in the first half are replaced with the corresponding positions in the second half, that is, for the sampling matrix B i The jth row of B ij ,have B ij =[b i,1 ,b i,2 ,…,b i,j-2 ,a i,j-1 ,b i,j ,…b i,D ],j=D+2,D+3,…,2D+1 Then we get a sampling matrix B of size (2D+2)N*D: <h2 style=";text-align:left;direction:ltr">B=[B1,B2,B3,…,B<h2 style=";text-align:left;direction:ltr"> N <h2 style=";text-align:left;direction:ltr"> ]<h2 style=";text-align:left;direction:ltr"> T S2-2, input matrix B into Catboost classifier to obtain output matrix C; S2-3. Calculate the Sobol first-order effect index, second-order effect index and total effect index of each feature according to matrix C; the calculation formula of Sobol first-order effect index is: Among them, S i Represents the i-th input variable X i The first-order effect index indicates the direct impact of a single input variable on the output variable; E(Y|X i ) is the expected value of the output variable given the i-th input variable; Y is the output variable; Var(·) represents the variance calculation; The calculation formula of Sobol second-order effect index is: Among them, S ij represents the second-order effect index between the i-th and j-th output variables, indicating the impact of the interaction of two input variables on the output variable; E(Y|X i ,X j ) is the expected value of the output variable under the condition of the i-th and j-th input variables; The calculation formula of Sobol total effect index is: in, is the total effect index of the ith input variable, indicating the total impact of a single input variable on the output variable; X ~i are all other input variables except the i-th input variable; E(Y|X ~i ) is the expected value of the output variable given all input variables except the ith input variable.
5. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 4 is characterized by: Sort the calculated Sobol first-order effect index, second-order effect index and total effect index, and output the ranking and degree of influence of the input variable on the output variable. The ranking is in descending order. The larger the effect index, the greater the influence of the input variable on the output variable. The grid operation data, temperature, coal price and time label of the predicted date are input into the trained CatBoost classifier, and the classification prediction result is output. The output result may be -1 or 1. When the result is 1, it means that the day-ahead electricity price is greater than the real-time electricity price. When the result is -1, it means that the day-ahead electricity price is less than the real-time electricity price.
6. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 1, characterized in that: The actual operation data of the power grid include the centralized load, Class A power output, Class B power output, local power output, external support power and interconnection line power; the coal price is the coal guidance price, among which Class A power refers to the power that has not participated in market transactions, and Class B power refers to the power that has participated in market transactions.
7. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 2 is characterized by: The time label includes the day of the week, month, day, hour and minute.
8. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 1 is characterized by: The training generates a CatBoost classifier by first training a base learner from the initial training set, and then adjusting the distribution of training samples according to the performance of the base learner so that the training samples that the previous base learner got wrong have greater weights in the subsequent period, and the weight range is [0,1]. Then, the next base learner is trained based on the adjusted sample distribution, and the process is repeated until the base learner data reaches a predetermined value T, and finally the T base learners are weightedly combined to obtain the final strong learner. The larger the T value is, the more complex the entire classifier is and the longer the training time is. In practice, the T value is usually determined by comprehensively considering the complexity of the problem and the available computing resources.
9. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 4, characterized in that: The sampling matrix B, the 1st to 13th input variables are: centralized load, Class A power output, Class B power output, local power output, external support power, interconnection line power, temperature, coal price, week sequence, month, day, hour, minute, and the output variable is the positive or negative price difference between the day-ahead market and the real-time market.
10. The method for predicting the day-ahead-real-time power market price spread considering global sensitivity analysis according to claim 4, characterized in that: The number of sampling samples N is selected as 2048 and the feature dimension D is selected as 13.
