A new energy participating in power market bidding method based on kernel optimization algorithm

Abstract

The application discloses a new energy participating in power market bidding method based on a kernel optimization algorithm, relates to the technical field of the fusion of power market operation management and machine learning algorithms, and comprises the following steps: S1, constructing a characteristic vector and preparing data; S2, performing infinite norm normalization processing on the characteristic vector; S3, calculating a weight through a Gaussian kernel function, and representing the similarity between a normalized characteristic vector at the next moment and a normalized characteristic vector of a historical sample through the weight; S4, constructing a single-step economic decision model, constructing a weighted experience cumulative distribution function CDF, and solving an optimal bid amount satisfying a target quantile; and S5, automatically executing bandwidth adjustment through a rolling verification mechanism to adapt to the seasonal drift characteristics of photovoltaic output power. The new energy participating in power market bidding method based on the kernel optimization algorithm can solve the fundamental mismatch problem that more accurate prediction does not equal to better decision in the prior art.

Description

Technical Field

[0001] This invention relates to the technical field of integrating power market operation management with machine learning algorithms, and in particular to a method for new energy sources to participate in power market bidding based on kernel optimization algorithms. Background Technology

[0002] my country's installed capacity of new energy sources has grown rapidly, becoming an important component of the power system. To promote the consumption of new energy, optimize resource allocation, and improve market efficiency, national policies are gradually shifting the focus of new energy purchases from guaranteed procurement to market-based transactions, requiring new energy power generation companies to participate in electricity market bidding through medium- and long-term transactions and spot markets. However, the output of new energy sources is characterized by significant intermittency, volatility, and uncertainty, and is greatly affected by weather conditions. This leads to multiple challenges in a market-based bidding environment, including price volatility risks and revenue uncertainty, necessitating efficient bidding strategies to maximize returns.

[0003] Under the current dual-price imbalance settlement mechanism in the electricity market, renewable energy generators are required to submit their bids 12-36 hours in advance, and deviations from actual output will incur hefty penalties. This decision-making problem is essentially a classic newsboy problem, requiring the determination of the optimal bid quantity under uncertain renewable energy output conditions to minimize the expected imbalance penalty cost.

[0004] Existing technologies primarily employ a "prediction-optimization" framework, belonging to the statistical basis learning paradigm, and can be divided into two categories: The first is the indirect method, which first performs parameterized estimation of the power output distribution (e.g., assuming a normal distribution and estimating the mean and variance using OLS), and then applies the critical quantile formula to solve for the bidding quantity. This type of method relies on strong distribution assumptions; under the high-dimensional characteristics and seasonal drift scenarios of photovoltaics, the prediction error in the first stage will be directly amplified to the decision-making stage, leading to large bidding deviations and high economic losses. The second is the featureless direct method, which directly uses the empirical quantiles of historical samples as the bidding quantity. This method is computationally simple, but completely ignores rich features such as numerical weather forecasts and lagged power output, resulting in low bidding accuracy and inability to adapt to complex photovoltaic scenarios.

[0005] The aforementioned existing technologies generally suffer from the following drawbacks: statistical loss (MSE, etc.) is inconsistent with decision loss; the impact of prediction error on the decision objective is asymmetric, nonlinear, and non-monotonic, and more accurate statistical prediction may not necessarily lead to better decision-making; parameterized methods are sensitive to distribution assumptions, and the non-normal characteristics of photovoltaic data lead to severe error propagation; featureless methods cannot utilize high-dimensional auxiliary information and have poor generalization ability; and computational efficiency is insufficient, making it difficult to meet the needs of real-time rolling bidding. Summary of the Invention

[0006] The purpose of this invention is to provide a method for new energy sources to participate in electricity market bidding based on kernel optimization algorithms, thereby solving the fundamental mismatch problem in existing technologies where more accurate predictions do not equate to better decisions.

[0007] To achieve the above objectives, this invention provides a method for new energy sources to participate in electricity market bidding based on a kernel optimization algorithm, comprising the following steps: S1. Construct feature vectors and prepare data; S2. Perform infinite norm normalization on the eigenvectors; S3. Calculate the weights using the Gaussian kernel function, and use the weights to characterize the similarity between the normalized feature vector at the next time step and the normalized feature vector of the historical samples. S4. Construct a single-step economic decision-making model, build a weighted empirical cumulative distribution function CDF, solve for the optimal reporting quantity that satisfies the target quantile, and directly use it for market reporting; S5. To adapt to the seasonal drift characteristics of photovoltaic power output, bandwidth optimization is automatically performed through a rolling verification mechanism.

[0008] Preferably, in S1: S11, construct 3D feature vector, front The dimensional feature is a real-time feature, the 3rd feature is a real-time feature. The characteristic is that it lags behind the actual photovoltaic output by one period. The feature dimension can be expanded according to actual needs. The expansion range of the feature dimension depends on the relative fit between the training sample size and the feature dimension. S12. Collect historical photovoltaic operation data and generate training set. The expression is: ; In the formula, Indicates the historical sample index; The training window length represents the total number of historical samples; Indicates the first Feature vectors of historical samples; Indicates the first The actual output of a historical sample; S13, Regarding the current decision-making moment Extract the feature vector at the next time step. This serves as a feature input for this bidding decision.

[0009] Preferably, in S2: S21, for the training set... The historical samples are normalized using the infinite norm, and the expression is: ; In the formula, Indicates the feature dimension index; Indicates the first The first historical sample 3D feature vector; Indicates the training set number 3D feature column vector; Indicates the first The first historical sample The eigenvectors after normalization of the eigenvectors; S22. Following the method in S21, for the feature vector at the next time step... Perform infinity norm normalization to obtain the normalized eigenvectors at the next time step. .

[0010] Preferably, in S3: S31, for the first training set... Weights are calculated from historical samples using a Gaussian kernel function. ; S32. Traverse all historical samples in the training set, complete the weight calculation for all samples, and form a weight set.

[0011] Preferably, the expression for calculating the weights in S31 is: ; In the formula, It is a hyperparameter representing the Gaussian kernel bandwidth, which will be automatically tuned later.

[0012] Preferably, in S4: S41, construct a single-step economic decision-making model through single-period unbalanced costs, without estimating any distribution, and directly minimize the kernel-weighted conditional costs; S42. Since the unbalanced cost of a single period is piecewise linearly convex with respect to the current reporting quantity to be decided, the optimal solution of the core surrogate function is equivalent to the weighted empirical quantile. The weighted empirical cumulative distribution function CDF is constructed and the optimal reporting quantity is found.

[0013] Preferably, the expression for the single-period unbalanced cost in S4 is: ; In the formula, Indicates insufficient penalty coefficient; This indicates the current number of reports awaiting a decision; Indicates actual historical contributions; This indicates the excess penalty coefficient.

[0014] Preferably, the expression for the core surrogate function of the single-step economic decision-making model in S4 is: ; In the formula, Indicates the feature vector based on the next time step. The kernel-weighted estimate of the conditional cost; Indicates the first The single-period imbalance cost of a historical sample.

[0015] Preferably, in S42: S421, sort the actual output of all historical samples in the training set from smallest to largest, and simultaneously sort the weights of all historical samples. S422. Accumulate and sort the weights to construct a weighted empirical cumulative distribution function; S423. Compare the accumulated weighted value with the target quantile until the first point where the accumulated weighted value is not less than the target quantile is reached. This point is the current decision point. The optimal reporting quantity, directly used for market reporting, is expressed as: ; In the formula, Indicates the current decision-making moment The optimal bid volume decision, that is, the optimal bid price ultimately used for market submission; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true and 0 when it is false. Describe the target quantile, satisfying .

[0016] Preferably, in S5: S51, Divide the validation set: After selecting the training set, the window length is... The data is used as a validation set to determine the candidate set of Gaussian kernel bandwidth; S52. Iterate through all candidate Gaussian kernel bandwidths on the validation set, calculate the average loss for each candidate Gaussian kernel bandwidth, and select the candidate Gaussian kernel bandwidth with the lowest average loss as the optimal bandwidth. .

[0017] Therefore, the present invention employs the above-mentioned method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm, which has the following beneficial effects: (1) By using a one-step kernel weighted optimization method, which is different from the two-step method, kernel regression is directly performed on the newsboy cost, thus avoiding the amplification of distribution estimation error in the optimization process; (2) By using Gaussian kernel similarity weighting, historical scenes are given higher weights, which can automatically capture the implicit pattern of "today's weather is like a certain day last year", significantly improving the accuracy of conditional quantiles, especially in scenarios with sudden changes in cloud cover and extreme weather. (3) The bandwidth optimization is automatically performed through the rolling verification mechanism, eliminating the need for manual experience to set the bandwidth. This allows the method of the present invention to adapt to the optimal parameters in different seasons and different power plants, with strong generalization ability, and eliminates the parameter sensitivity defect of traditional methods. (4) By using a sorting algorithm to solve the problem, only one sorting and accumulation is required, which reduces the computational complexity. The single point takes only 0.00573 seconds, which meets the needs of a large number of new energy power stations for parallel bidding and reduces server costs.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of a method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to the present invention; Figure 2 This is a block diagram illustrating the kernel optimization solution principle of Embodiment 1 of the present invention; Figure 3 This is a comparison curve of the actual output and the optimal KO price of the test set in Embodiment 1 of the present invention; Figure 4 This is the bandwidth tuning curve of the verification set in Embodiment 1 of the present invention; Figure 5 This is an experimental result diagram of Embodiment 2 of the present invention. Detailed Implementation

[0020] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] Example 1 Please see Figure 1-4 A kernel optimization algorithm-based method for new energy sources to participate in electricity market bidding, referred to as the KO method, includes the following steps: S1. Construct feature vectors and prepare data; S11. Construct a 13-dimensional feature vector. The first 12 dimensions include real-time features such as photovoltaic meteorology, geography, real-time irradiance, temperature, and cloud cover. The 13th dimension is the actual photovoltaic output with a lag of 1 period. The feature dimensions can be expanded according to actual needs. The expansion range of the feature dimensions depends on the relative fit between the training sample size and the feature dimensions. S12. Collect historical photovoltaic operation data and generate training set. The expression is: ; In the formula, Indicates the historical sample index; The training window length represents the total number of historical samples; Indicates the first Feature vectors of historical samples; Indicates the first The actual output of a historical sample; S13, Regarding the current decision-making moment Extract the feature vector at the next time step. This serves as a feature input for this bidding decision.

[0022] S2. To eliminate the dimensional differences between different features and ensure the stability of the distance metric in the high-dimensional feature space, the feature vectors are normalized using the infinite norm. S21, regarding the training set... The historical samples are normalized using the infinite norm, and the expression is: ; In the formula, Indicates the feature dimension index; Indicates the first The first historical sample 3D feature vector; Indicates the training set number 3D feature column vector; Indicates the first The first historical sample The eigenvectors after normalization of the eigenvectors; S22. Following the method in S21, for the feature vector at the next time step... Perform infinity norm normalization to obtain the normalized eigenvectors at the next time step. .

[0023] S3. Calculate the weights using the Gaussian kernel function. The weights represent the similarity between the normalized feature vector at the next time step and the normalized feature vector of the historical samples. The higher the similarity, the greater the weight of the corresponding historical sample, providing a basis for subsequent cost weighting. S31, targeting the training set Weights are calculated from historical samples using a Gaussian kernel function. The expression is: ; In the formula, It is a hyperparameter representing the Gaussian kernel bandwidth, which will be automatically tuned later. The expression for weight calculation is derived from the Gaussian kernel expression and the expression incorporating kernel bandwidth; the Gaussian kernel expression is: ; In the formula, Represents the standard Gaussian kernel function; The expression for incorporating kernel bandwidth is: ; In the formula, Indicates the first The kernel weight function for each historical sample; S32. Traverse all historical samples in the training set, complete the weight calculation for all samples, and form a weight set.

[0024] S4. Construct a single-step economic decision-making model, build a weighted empirical cumulative distribution function CDF, solve for the optimal reporting quantity that satisfies the target quantile, and directly use it for market reporting; S41. Construct a single-step economic decision-making model, without estimating any distribution, directly minimizing the kernel-weighted conditional cost. The expression for the core surrogate function is: ; In the formula, Indicates the feature vector based on the next time step. The kernel-weighted estimate of the conditional cost; This indicates the current number of reports awaiting a decision; Indicates the first The single-period imbalance cost of a historical sample; Single-period unbalanced costs satisfy: ; In the formula, Indicates insufficient penalty coefficient; Indicates actual historical contributions; Indicates the excess penalty coefficient; S42. Since the single-period unbalanced cost is piecewise linearly convex with respect to the current reporting quantity to be decided, the optimal solution of the core surrogate function is equivalent to the weighted empirical quantile. Construct the weighted empirical cumulative distribution function CDF and find the optimal reporting quantity. S421. Sort the actual output of all historical samples in the training set from smallest to largest, and simultaneously sort the weights of all historical samples. S422. Accumulate and sort the weights to construct a weighted empirical cumulative distribution function; S423. Compare the accumulated weighted value with the target quantile until the first historical sample with an accumulated weighted value not less than the target quantile is identified as the output point. The actual output value of this historical sample is the optimal output at the current decision moment, i.e., this point is the optimal output at the current decision moment. The optimal reporting quantity, directly used for market reporting, is expressed as: ; In the formula, Indicates the current decision-making moment The optimal bid volume decision, that is, the optimal bid price ultimately used for market submission; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true and 0 when it is false. Describe the target quantile, satisfying .

[0025] S5. To adapt to the seasonal drift characteristics of photovoltaic power output, bandwidth optimization is automatically performed through a rolling verification mechanism. S51. Divide the validation set: After selecting the training set, the window length is... The data is used as a validation set to determine the candidate set of Gaussian kernel bandwidth; S52. Iterate through all candidate Gaussian kernel bandwidths on the validation set, calculate the average loss for each candidate Gaussian kernel bandwidth, and select the candidate Gaussian kernel bandwidth with the lowest average loss as the optimal bandwidth. .

[0026] Experimental verification To verify the practical effect of this invention, simulation experiments were conducted using the GEFCom2014 photovoltaic dataset (1500MW installed capacity). A 13-dimensional feature vector was constructed. The first 12 dimensions included real-time features such as photovoltaic meteorology, geography, real-time irradiance, temperature, and cloud cover (from the GEFCom2014 photovoltaic dataset), and the 13th dimension was the actual photovoltaic output with a lag of one period. The training set consisted of 10 consecutive weeks from October 1, 2012 (a total of 1680 hours, with a resolution of 1 hour); the validation set consisted of the last 840 hours of the training set, used for automatic bandwidth optimization; the test set consisted of 200 consecutive hours, used for performance evaluation; cost parameters were set, including... , and The method of the present invention was compared with the featureless direct method (SAA) and the two-step method (SEO); the experimental results are shown in Table 1; compared with the traditional two-step method, the KO method of the present invention can reduce the average cost by 19.78% in the photovoltaic bidding scenario, and the calculation time for a single point is only 0.00573 seconds, which fully meets the requirements of real-time bidding in the power market.

[0027] Table 1: Performance Comparison of Different Methods

[0028] Example 2 Consider the k-nearest neighbor (KNN) weighted quantile method, which uses a fixed set of k nearest neighbor samples to replace the Gaussian kernel and directly sorts them to calculate the quantiles. The rest of the process is the same as the method in Example 1, and will not be described in detail here.

[0029] Specific implementation method: Select k nearest neighbor samples, use the inverse of Euclidean distance as weight, directly sort the historical output and construct a weighted empirical cumulative distribution function. No kernel function calculation is required. It only includes two steps: distance sorting and weighted accumulation.

[0030] Please see Figure 5The method of Example 2 is applied to small and medium-sized distributed photovoltaic power stations, with an installed capacity of... , The experimental results are shown in Table 2.

[0031] Table 2: Experimental results of the k-nearest neighbor (kNN) weighted quantile method

[0032] Example 3 Replace the kernel function with a uniform kernel, Epanechnikov kernel, or other positive definite kernel, and only modify the weight formula. Perform the same procedure as in Example 1, and will not be elaborated further.

[0033] If the uniform kernel is used as the kernel function, then the weight expression is: ; In the formula, The indicator function represents the uniform kernel function. It takes the value 1 when the condition in parentheses is true, and takes the value 0 otherwise. Indicates uniform kernel bandwidth; If the Epanechnikov kernel is used as the kernel function, the weight expression is: ; In the formula, This indicates the Epanechnikov kernel bandwidth; Other positive definite kernels (such as trigonometric kernels and cosine kernels) can be replaced in the same way.

[0034] Example 4 The feature vector is expanded by adding power output, temperature and radiance quadratic cross terms for the same time period with a 24-hour lag, and real-time load characteristics of the power grid. The expanded feature vector is normalized to the infinite norm, and the kernel weight calculation formula remains unchanged. Only the feature dimension is increased. Other contents are the same as those in Example 1, and will not be described in detail.

[0035] Therefore, this invention adopts a kernel optimization algorithm-based method for new energy participation in electricity market bidding. By using a one-step kernel-weighted optimization method, distinct from the two-step method, it directly performs kernel regression on the newsboy cost, avoiding the amplification of distribution estimation errors in the optimization process. Through Gaussian kernel similarity weighting, historical scenarios are given higher weight, automatically capturing the implicit pattern of "today's weather is like a certain day last year," significantly improving the accuracy of conditional quantiles, especially under sudden cloud cover changes and extreme weather scenarios. The rolling verification mechanism automatically performs bandwidth optimization, eliminating the need for manual bandwidth settings. This allows the method to adapt to optimal parameters in different seasons and at different power plants, exhibiting strong generalization ability and eliminating the parameter sensitivity defects of traditional methods. The sorting algorithm solution requires only one sorting and accumulation, reducing computational complexity. The single-point time is only 0.00573 seconds, meeting the parallel bidding needs of massive new energy power plants while reducing server costs.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm, characterized in that, Includes the following steps: S1. Construct feature vectors and prepare data; S2. Perform infinite norm normalization on the eigenvectors; S3. Calculate the weights using the Gaussian kernel function, and use the weights to characterize the similarity between the normalized feature vector at the next time step and the normalized feature vector of the historical samples. S4. Construct a single-step economic decision-making model, build a weighted empirical cumulative distribution function CDF, and solve for the optimal rate that satisfies the target quantile. S5. To adapt to the seasonal drift characteristics of photovoltaic power output, bandwidth optimization is automatically performed through a rolling verification mechanism.

2. The method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 1, characterized in that, In S1: S11, Construction 3D feature vector, front The dimensional feature is a real-time feature, the 3rd feature is a real-time feature. The characteristic is that it lags behind the actual photovoltaic output by one period. Indicates the feature dimension; S12. Collect historical photovoltaic operation data and generate training set. The expression is: ； In the formula, Indicates the historical sample index; This represents the total number of historical samples; Indicates the first Feature vectors of historical samples; Indicates the first The actual output of a historical sample; S13, Regarding the current decision-making moment Extract the feature vector at the next time step. .

3. The method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 2, characterized in that, In S2: S21, regarding the training set... The historical samples are normalized using the infinite norm, and the expression is: ； In the formula, Indicates the feature dimension index; Indicates the first The first historical sample 3D feature vector; Indicates the training set number 3D feature column vector; Indicates the first The first historical sample The eigenvectors after normalization of the eigenvectors; S22. Following the method in S21, for the feature vector at the next time step... Perform infinity norm normalization to obtain the normalized eigenvectors at the next time step. .

4. The method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 3, characterized in that, In S3: S31, targeting the training set Weights are calculated from historical samples using a Gaussian kernel function. ; S32. Traverse all historical samples in the training set, complete the weight calculation for all samples, and form a weight set.

5. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 4, characterized in that, The expression for calculating the weights in S31 is: ； In the formula, It is a hyperparameter representing the Gaussian kernel bandwidth.

6. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 5, characterized in that, In S4: S41. Construct a single-step economic decision-making model using single-period unbalanced costs; S42. Since the single-period unbalanced cost is piecewise linearly convex with respect to the current decision-making quantity, the optimal solution of the core surrogate function is equivalent to the weighted empirical quantile. The weighted empirical cumulative distribution function CDF is constructed and the optimal quantity is found.

7. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 6, characterized in that, S4 single-period imbalance cost The expression is: ； In the formula, Indicates insufficient penalty coefficient; This indicates the current number of reports awaiting a decision; Indicates actual historical contributions; This indicates the excess penalty coefficient.

8. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 7, characterized in that, The expression for the core surrogate function in the single-step economic decision-making model of S4 is: ； In the formula, Indicates the feature vector based on the next time step. The kernel-weighted estimate of the conditional cost; Indicates the first The single-period imbalance cost of a historical sample.

9. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 8, characterized in that, In S42: S421. Sort the actual output of all historical samples in the training set from smallest to largest, and simultaneously sort the weights of all historical samples. S422. Accumulate and sort the weights to construct a weighted empirical cumulative distribution function; S423. Compare the accumulated weighted value with the target quantile until the first point where the accumulated weighted value is not less than the target quantile is reached. This point is the current decision point. The optimal reporting quantity, directly used for market reporting, is expressed as: ； In the formula, Indicates the current decision-making moment The optimal reporting decision; Indicates an indicator function; Describe the target quantile, satisfying .

10. A method for new energy sources to participate in electricity market bidding based on kernel optimization algorithm according to claim 9, characterized in that, In S5: S51. Divide the validation set: After selecting the training set, the window length is... The data is used as a validation set to determine the candidate set of Gaussian kernel bandwidth; S52. Iterate through all candidate Gaussian kernel bandwidths on the validation set, calculate the average loss for each candidate Gaussian kernel bandwidth, and select the candidate Gaussian kernel bandwidth with the lowest average loss as the optimal bandwidth. .

Images