Power supply optimization model and method for converter station

Abstract

The invention provides a power supply optimization model and method for a converter station. The method comprises the steps that firstly, a converter station energy consumption characteristic analysis and electric quantity prediction model is constructed, and the converter station energy consumption characteristic analysis and electric quantity prediction model comprises user converter station energy consumption characteristic analysis based on wavelet transformation and an optimization fast density peak value clustering algorithm and a converter station electric quantity prediction model based on an XGBoost algorithm; then, constructing a wind power and photovoltaic novel power supply output prediction model, and adopting a wavelet decomposition-bidirectional long-short term memory network based on an Attention mechanism as a prediction model; and finally, in combination with the electric quantity prediction model and the novel power supply output prediction model, constructing a power supply combination optimization design model for the converter station. Through the model, the power optimization capability and the use efficiency of the converter station can be improved, electric energy resources are managed and reasonably utilized, and the economical efficiency and the safety of a power system are improved. The power supply optimization model and method for the converter station are innovative and have practical value.

Description

Technical Field

[0001] The present invention belongs to the field of grid converter stations, and more specifically relates to an optimized model and method for the station service power supply of a converter station. Background Art

[0002] In the power system, converter stations are an important part of the power transmission system. With the rapid development and application of new energy technologies, such as wind energy, photovoltaics, etc., the power consumption of new types of power sources is increasing, and the station service power supply of converter stations is gradually changing. How to combine new energy power sources and traditional power sources to optimize the design of the station service power supply of converter stations is an important issue faced by the current power system.

[0003] Specifically, modern power systems can already use big data technology for in-depth learning and prediction to achieve accurate prediction of power demand.

[0004] First, regarding the characteristics of electrical load, it mainly includes research fields such as data preprocessing, outlier detection, load clustering analysis, and index evaluation. These are all achieved by analyzing and mining a large amount of data to improve the quality and efficiency of power services. Among them, load curve clustering, as an important analysis method, commonly used techniques include K-means, fuzzy C-means (FCM), hierarchical clustering, self-organizing map (SOM), and density clustering (DSCAN), as well as other algorithms such as SVM, GMM, etc. The load curve is clustered in one or more ways to better understand and predict the power demand of users.

[0005] Next is the research on user load clustering. Industry experts have developed a series of complex clustering algorithms, such as the adaptive K-means algorithm and improved methods based on the K-means algorithm, and have conducted a large number of empirical analyses using these clustering algorithms. Without a doubt, in addition to improving and optimizing existing clustering algorithms, reducing the time complexity of the algorithms is also one of the urgent problems to be solved.

[0006] In the research field of the load characteristics of large industrial users, load forecasting and off-peak power consumption management are two main application scenarios. Among them, load forecasting can improve the accuracy of forecasting by analyzing load characteristics; off-peak power consumption is to implement off-peak power consumption strategies based on load characteristics, thereby balancing the power load, reducing the pressure on the power grid, and ensuring the reliable operation of the power grid.

[0007] In summary, power systems at home and abroad have entered a new stage. By using big data technology, power demand forecasting is more accurate, and power services have become more efficient and intelligent. However, at the same time, we still need to find more accurate and practical forecasting models and algorithms to better adapt to future changes in power demand. Summary of the Invention

[0008] The present invention provides an optimized model and method for the station service power supply of a converter station, including an energy consumption characteristic analysis and power consumption prediction model for the converter station, an output power prediction model for new energy sources such as wind power and photovoltaic power, and an optimized design model for the combination of the station service power supply of the converter station. By deeply analyzing the power consumption characteristics of the converter station and combining the output power prediction of new energy sources, the power supply combination is scientifically configured, so as to optimize the operation of the converter station and improve the economic benefits of the power system.

[0009] To achieve the above object, the present invention is implemented by the following technical solutions: The method includes:

[0010] Construct an energy consumption characteristic analysis and power consumption prediction model for the converter station, and adopt an energy consumption characteristic analysis of the user converter station based on wavelet transform and optimized fast density peak clustering algorithm, as well as a power consumption prediction model for the converter station based on the XGBoost algorithm;

[0011] Construct an output power prediction model for new energy sources such as wind power and photovoltaic power, and adopt a wavelet decomposition - bidirectional long short - term memory network based on the Attention mechanism as the prediction model;

[0012] Combine the power consumption prediction model and the output power prediction model of new energy sources to construct an optimized design model for the combination of the station service power supply of the converter station.

[0013] In one solution, the construction of the energy consumption characteristic analysis and power consumption prediction model for the converter station includes:

[0014] Power consumption prediction of the converter station based on the XGBoost algorithm;

[0015] Energy consumption characteristic analysis of the user converter station based on wavelet transform and optimized fast density peak clustering algorithm, design an optimized fast density peak clustering algorithm based on wavelet transform, which is an improved algorithm by combining wavelet transform, KNN algorithm, fast density peak clustering algorithm and clustering center automatic selection strategy.

[0016] In one solution, the energy consumption characteristic analysis of the user converter station by the wavelet transform and optimized fast density peak clustering algorithm is respectively wavelet transform, clustering analysis and reconstruction analysis;

[0017] For the wavelet transform, the sample data is subjected to wavelet transform and decomposed into spectral components with different time scales. The spectral components obtained with different decomposition levels are also different. Each time of decomposition, the time scale is reduced by half. After decomposing n layers, the detail components D, D…, D, and the approximation component A can be obtained;

[0018] For the above-mentioned clustering analysis, the spectral components D, D…, D and 4 after wavelet decomposition are used as inputs, and clustering analysis is performed separately; the K nearest neighbors of each sample point are found, and the local density p and KNN distance 8 of each sample point are calculated using the optimized fast density peak clustering algorithm, so as to obtain the decision variable. Then, the clustering centers of different spectral components are obtained by using the automatic clustering center selection strategy, and finally, all sample points are labeled with their respective categories.

[0019] For the above-mentioned reconstruction analysis, the clustering centers of each spectral component are reconstructed. The clustering centers of different spectral components are denoted as TLP and TLP, and the typical load characteristic curve of the original data is TLP; by analyzing the reconstructed typical load curve, the electricity consumption characteristics of users can be understood.

[0020] In one solution, the construction of the output prediction model for new types of wind power and photovoltaic power sources includes:

[0021] Cleaning of wind, light, and water monitoring data, including: abnormal data detection, autocorrelation and partial autocorrelation analysis, and correlation analysis between power generation and meteorological factors;

[0022] For the output prediction model of wind, light, and water renewable resources power generation, the renewable energy power prediction model uses historical power data and factors affecting the change of power values as input parameters, and the power value to be predicted as the output. Then, the power prediction problem is transformed into a mathematical problem of finding the mapping relationship between the input variable X and the output variable Y. A wavelet decomposition-bi-directional long short-term memory network based on the Attention mechanism is used as the prediction model.

[0023] In one solution, the optimal design model for the station service power supply combination of the converter station includes:

[0024] Multi-objective resource screening model, which decomposes the problem into the composition method of coalition members and task allocation, and uses modern portfolio theory to transform the response power allocation problem into the allocation weights of risky and risk-free assets;

[0025] Risk assessment of the resource combination of the converter station, which is analyzed according to the mean-variance analysis method and the MPT efficient frontier model in the MPT theory;

[0026] Analysis of the output characteristics of the station service power supply. According to the control authority of the converter station over resource equipment, the resources are divided into two categories: controllable and uncontrollable; in the contract signed between the converter station and the user, if the user does not entrust the control authority of the terminal equipment to the converter station, it is regarded as a demand response user. The response power provided by such users is uncertain and regarded as a risky asset. Such users are mostly large load users with various adjustable electricity consumption characteristics inside and their own energy management systems.

[0027] Risk - free asset response characteristic model. In the contract signed between the converter station and the user, if the user entrusts the control right of the terminal equipment to the converter station, the response power provided by such users is regarded as a risk - free asset; the response power provided by such users is essentially a risk - free asset with σ = 0 and ρ = 0, avoiding the revenue risk caused by the user's response willingness. Considering that the technical characteristics of the terminal equipment entrusted to the converter station are different, the user - side interactive resources are divided into two categories: flexible controllable load and DG.

[0028] In one solution, the above - mentioned converter station station - use power combination optimization design model further includes:

[0029] Converter station resource optimization combination model. The optimization combination model is established as follows:

[0030] max:E(R p )=C in -E(C cost )

[0031]

[0032] In the model, with the goal of maximizing the expected revenue, in the objective function, E(R p ) represents the expected revenue of the converter station after combination; C in represents the expected income of the converter station. After the market clearing determines the bidding curve P aim,t , the expected revenue C in =(C r -C w )P aim,t Δt is a determined constant value. The profit margin comes from the difference between the retail electricity price C r and the wholesale electricity price C w in the electricity market. E(C cost ) is the expected cost, which consists of three parts. The expression is as follows:

[0033]

[0034] In the formula, f1 is the incentive cost required to encourage users to change the output plan to provide response power; f2 is the uncertainty cost generated by the deviation between the actual power and the target power of the converter station, corresponding to the penalty cost from the electricity market during the settlement stage, and c er is the unit deviation penalty price; f3 is the access cost caused to communication and computing resources required to maintain users' participation in the interaction of the converter station. C i,j,k is the unit response incentive cost of the interactive resource, that is, the internal trading electricity price; λ is the access cost required for a single resource to participate in regulation; V i,j,k is a 0 / 1 variable, and V i,j,k =1 indicates that this resource participates in the response. In the constraint conditions, h(Pt i,j,k ) ≥ 0 represents the risk and utility constraints for the optimal combination scheme, u(P t i,j,k ) ≥ 0 represents the resource response characteristic constraints shown in formula), g(P t i,j,k ) = 0 is the equality constraint representing the aggregated external characteristics of the converter station, s(P t i,j,k ) ≥ 0 represents the application constraints. In the formula, the portfolio weights of various assets are expressed as

[0035] In a scheme, for the solution of the optimization design model of the converter station station service power supply combination, an optimization heuristic solution algorithm is adopted.

[0036] In a scheme, the specific process of the output power prediction model for new types of power sources such as wind power and photovoltaic power is as follows:

[0037] Step1: Preprocess the obtained input data. In order to eliminate the influence of dimensions, eliminate the dimension error, and accelerate the training process, the input training data is normalized and converted into scalar values, and the output power values are obtained;

[0038] Step2: Use the wavelet decomposition method to perform wavelet decomposition and single-branch reconstruction on the input time series data, realize the optimal extraction of trend information, filter the fluctuation information, separate the noise, and obtain the final input sequence of the model;

[0039] Step3: Initialize the BiLSTM network based on the Attention mechanism, input the decomposed dataset training data into the network respectively, and use the training samples to train the network. The mean square error is used as the loss function, and the Adam optimization algorithm is used for network training to update the weights and obtain the prediction model;

[0040] Use the trained neural network model to predict each subsequence decomposed for the prediction period respectively, add the predicted subsequences, verify the prediction accuracy, and obtain the final predicted wind power and photovoltaic power data.

[0041] Advantages of the present invention:

[0042] The optimized model and method for the station power supply of the converter station of the present invention utilize wavelet transform and the optimized fast density peak clustering algorithm for the analysis of the energy consumption characteristics of the user's converter station, as well as the power consumption prediction model of the converter station based on the XGBoost algorithm. Further, the wavelet decomposition-bi-directional long short-term memory network based on the Attention mechanism is adopted as the output power prediction model for new power sources such as wind power and photovoltaic power. By combining the power consumption prediction model and the output power prediction model of the new power sources, an optimized design model for the combined station power supply of the converter station is constructed. This model and method can improve the power supply optimization performance of the converter station, enhance the efficiency of power supply utilization, effectively manage and rationally utilize electric energy resources, and improve the economy and security of the power system. Description of the Drawings

[0043] Figure 1 It is the energy consumption characteristic analysis diagram of the converter station of the present invention;

[0044] Figure 2 Schematic diagram of the Attention mechanism;

[0045] Figure 3 Frame diagram of the distributed renewable energy power generation prediction;

[0046] Figure 4 It is the schematic diagram of the load comparison before and after scheduling;

[0047] Figure 5 Schematic diagram of the improved boundary crossing aggregation method;

[0048] Figure 6 Flow chart of MOEA / D;

[0049] Figure 7 Schematic diagram of the CPS update perturbation mechanism. Detailed Implementation Manner

[0050] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. The typical embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0051] In order to effectively optimize the energy consumption of converter stations, improve energy utilization efficiency and significantly reduce the inaccuracy of electricity quantity prediction, thus ensuring the safe and stable operation of the power system and the improvement of power quality, a series of comprehensive steps and strategies have been carried out. First, the complexity and randomness of converter station energy consumption are analyzed, and it is recognized that these characteristics are affected by multiple factors such as economic environment, climate change, holiday arrangements and equipment status. In particular, the particularity and stability requirements of the DC transmission system are considered. On this basis, an electricity quantity prediction model is constructed, breaking through the limitations of traditional linear regression algorithms. The XGBoost algorithm is innovatively introduced. With its strong adaptability based on the boosting tree additive model and CART regression tree, it realizes high-precision prediction of complex power data sets and significantly improves the prediction effect. Furthermore, a comprehensive optimization model for renewable energy such as wind, light, water and energy storage is proposed, aiming to improve the energy utilization efficiency of converter stations through resource integration, reduce dependence on traditional energy, reduce pollutant emissions and protect the ecological environment. On the premise of ensuring system stability, the system stability of converter stations is enhanced by optimizing measures such as water abandonment strategies, ensuring the continuity and reliability of power transmission, and supplemented by reasonable energy scheduling and management to further reduce operating costs. At the technical implementation level, a method for analyzing the energy consumption characteristics of user converter stations based on wavelet transform and optimized fast density peak clustering algorithm, as well as the specific application path of the XGBoost algorithm in electricity quantity prediction, are elaborated in detail.

[0052] An optimization model and method for the station service power supply of a converter station, specifically including:

[0053] Construct an energy consumption characteristic analysis and electricity quantity prediction model for the converter station.

[0054] Electricity consumption prediction of converter stations based on the XGBoost algorithm

[0055] The weak classifier used by the XGBoost algorithm is the CART decision tree. This decision tree model is a method for the conditional probability distribution of the output variable Y under the condition of the given input random variable X... The basic model of the XGBoost algorithm is to construct an additive model composed of K CART regression trees with different weights. The expected goal is to make the predicted value of the entire tree group model as close as possible to the true value while having as large a generalization ability as possible. In each round of fitting the data, the leaf nodes of each weak learner (i.e., CART decision tree) are scored, and the weights of the weak learners are adjusted according to the fitting situation.

[0056] The form of its model is as follows:

[0057] Given the data set D = {(xi, yi)}, XGBoost uses K CART decision trees to learn the data set,

[0058] Then the model function is as shown in the following formula:

[0059]

[0060] In the above formula, f k (x) represents a CART regression tree, and H represents the hypothesis space, as shown in the following formula:

[0061] H = {f(X) = w q (X)}

[0062] In the above formula, q(X) represents that the sample X is assigned to a certain leaf node, and w represents the score of that leaf node (i.e., scoring the sample), so represents the predicted value of the CART regression tree for this sample.

[0063] 2) XGBoost objective function and its learning strategy

[0064] XGBoost is composed of K CART decision trees. Through multiple rounds of iteration on the data, it adjusts the weights between the K decision trees and scores the leaf nodes of each decision tree, which is an additive model. Based on this, the objective function of the XGBoost model is shown in the following formula:

[0065] Obj(Θ) = L(Θ) + Ω(Θ)

[0066] Among them, L(Θ) is the loss function, and Ω(Θ) is the regularization term.

[0067] What XGBoost internally adopts is as shown in the following formula:

[0068]

[0069] Among them, T is the number of leaf nodes, and w is the score of the leaf node. In terms of optimizing the loss function, XGBoost performs a second-order Taylor expansion on the loss function. After the t-th iteration, the prediction result of the model is equal to the sum of the model predictions of the previous t - 1 times plus the prediction of the t-th tree, as shown in the following formula:

[0070] ∧(t) ∧(t - 1)

[0071] y i = y i + f t (x i )

[0072] Then, combined with the previous regularization term function, the objective function at this time can be written as shown in the following formula:

[0073]

[0074] In this formula, the parameters y of the first t1 trees have all been obtained, and at this time, only the t-th tree needs to be learned. Then, the function is expanded in a second-order Taylor series at, and the result is shown in the following formula:

[0075]

[0076] In this formula Next, the constant term in the formula is removed, and Ω(f i ) is converted according to the format of the formula and substituted into the above formula, and the final objective function obtained is the following formula:

[0077]

[0078] The left term in the above formula is the accumulation of samples, and the right two terms are the accumulation of the penalty terms for leaf nodes. To unify these two terms, we further define the sample set on leaf node i as I j ={i|q(x i )=j}, then the above formula can be corrected as the following formula:

[0079]

[0080] When the tree structure is determined, to minimize the objective function, we can set the derivative of the above formula to 0, and the optimal prediction score for each leaf node is obtained as shown in the following formula:

[0081]

[0082] Then the minimum loss is shown in the following formula:

[0083]

[0084] After obtaining the minimum loss function and the formula for the optimal prediction score of leaf nodes, only by determining the tree structure can the above results be obtained. In the XGBoost algorithm, the method adopted is the greedy method. That is, each time a leaf node is tried to be split, and the gain before and after the split is calculated, and finally the split method with the largest gain is selected. The gain before and after

[0085] The gain calculation is based on the term in the minimum loss formula to calculate the loss contribution of each leaf node. Because there is a in front, the larger this term is, the smaller the loss is.

[0086] 3) The importance method built into XGBoost itself belongs to the embedding method. The specific steps are as follows:

[0087] Step 1. Import the dataset and classify the data according to the enterprise id.

[0088] Step 2. Call the xgboost module in the third-party library. First, use the importance function to screen features. Taking the company with enterprise ID = 1 as an example, after using the importance function, a contribution degree table of features to the results will be obtained.

[0089] (2) Energy consumption feature analysis of user converter stations based on wavelet transform and optimized fast density peak clustering algorithm

[0090] 1) Overall framework analysis of energy consumption features

[0091] An improved algorithm based on wavelet transform, KNN algorithm, fast density peak clustering algorithm and automatic selection strategy of clustering center is designed by combining wavelet transform, KNN algorithm, fast density peak clustering algorithm and automatic selection strategy of clustering center. Before clustering, wavelet transform is used to decompose the data at multiple scales. When using this method to analyze user load characteristics, it mainly includes three parts. The first part is wavelet transform. First, decompose the user load data to obtain the spectral components at different time scales, including detail components and approximation components; the first part is clustering analysis. Cluster analysis is performed on the spectral components at different time scales respectively, and the clustering centers of each spectral component are obtained using the optimized fast density peak clustering algorithm; the third part is reconstruction analysis. Curve reconstruction is performed on the clustering centers of each spectral component to obtain the original typical load characteristic curve.

[0092] In the process of clustering analysis of the user load curve, our team improves the fast density peak clustering algorithm, uses the fast density peak clustering algorithm based on KNN for clustering to improve the clustering accuracy, and combines the automatic selection strategy of clustering center to select the clustering center to avoid artificial determination of the clustering center. The energy consumption feature analysis process of the converter station is as follows Figure 1 as shown.

[0093] 2) Basic theory of the algorithm

[0094] The energy consumption feature analysis mainly includes three parts: wavelet transform, clustering analysis and reconstruction analysis. After three steps, the typical load characteristic curve of the user can be obtained, and a small number of typical load characteristic curves are used to represent a large amount of user load data. The following will introduce the user load characteristic analysis process of the optimized fast density peak clustering algorithm based on wavelet transform from these three parts.

[0095] (1) Wavelet transform performs wavelet transform on the sample data, decomposes it into spectral components with different time scales, and different spectral components are obtained for different decomposition layers. Each time it is decomposed, the time scale is reduced by half. After decomposing n layers, the detail components D, D…, D, and the approximation component A, can be obtained. The specific number of decomposition layers needs to be determined according to the sampling frequency of the sample and experimental experience.

[0096] (2) Cluster analysis

[0097] Taking the spectral components D, D…, D and 4 after wavelet decomposition as inputs, perform cluster analysis respectively. Find the K nearest neighbors of each sample point, use the optimized fast density peak clustering algorithm to calculate the local density p and the KNN distance δ of each sample point, so as to obtain the decision variable, and then use the automatic selection strategy of the cluster center to obtain the cluster centers of different spectral components. Finally, label the categories to which all sample points belong.

[0098] (3) Reconstruction analysis

[0099] The cluster centers obtained by clustering different spectral components are only the load characteristic curves at this scale. If the load characteristics of the original data are to be obtained, the clustering of each spectral component needs to be carefully reconstructed. The cluster centers of different spectral components are denoted as TLP and TLP, and the typical load characteristic curve of the original data is TLP:

[0100]

[0101] By analyzing the reconstructed typical load curve, understanding the user's electricity consumption characteristics, effective suggestions for peak-shaving electricity consumption can be put forward. Referring to the user's load characteristics, time-of-use electricity prices can be designed specifically, and personalized electricity consumption suggestions can be reasonably put forward for users to obtain greater economic benefits.

[0102] Construct a prediction model for the output of new power sources such as wind power and photovoltaic power

[0103] (1) Autocorrelation and partial autocorrelation analysis

[0104] Since the prediction of wind power and photovoltaic power is based on historical power data information such as wind and light, it is necessary to perform autocorrelation analysis and partial autocorrelation analysis on wind and light power data to judge the stationarity of time series data, and at the same time obtain the long-term trend (Trend), seasonality (Seasonality), and residuals (Residuals) of the data.

[0105] The autocorrelation function (AutocorrelationFunction, ACF) reflects the correlation between the values of the same sequence at different time series, that is, the correlation degree between the current value of the sequence and its historical values. For the time series X = {x t , t = 1, 2,......., n}, its autocovariance function of lag k order is:

[0106]

[0107] where is the time series {x tThe average value of the elements in. Let k = 0, we get:

[0108]

[0109] Therefore, the autocorrelation coefficient of the time series {x t} is:

[0110]

[0111] The partial autocorrelation function (PACF) is a correlation measure that describes the influence of the time series {x t-k} on {x t}, while excluding the interference of the middle k - 1 random variables, that is, the interference of {x t-1 , x t-2 ,......, x t-k-1}. Its expression is:

[0112]

[0113] (2) Correlation analysis between power generation and meteorological factors

[0114] Since there is a large correlation between renewable new energy power generation such as wind power, photovoltaic power, and hydropower and meteorological factors, it is necessary to conduct a correlation analysis between power generation and meteorological factors. Based on the Pearson correlation coefficient and the maximum information coefficient correlation analysis method, the renewable new energy such as wind power, photovoltaic power, and hydropower is analyzed to comprehensively analyze the correlation between each influencing factor and power generation.

[0115] For the variable sequences X = {x i , i = 1, 2, ……, n} and Y = {y i , i = 1, 2,......, n}, the calculation formula of the Pearson correlation coefficient is as follows:

[0116]

[0117] Where cov(X, Y) is the covariance between variables X and Y, σ X and σ Y are the standard deviations between variables X and Y. If the Pearson coefficient r XY of variables X and Y is greater than 0, then the two variables are positively correlated; if the Pearson coefficient r XY of variables X and Y is less than 0, then the two variables are negatively correlated. The larger |r XY | is, the higher the degree of correlation between variables X and Y.

[0118] The calculation of the maximum information coefficient is based on mutual information and the grid partitioning method. For the variable sequence X = {xi , i = 1, 2,......, n} and Y = {y i , i = 1, 2,......, n}, the formula for calculating their mutual information MI(X, Y) is as follows:

[0119]

[0120] where p(x) and p(y) are the marginal probability densities of variables X and Y, and p(x, y) is the density of the joint probability of variables X and Y.

[0121] Let the variables X and Y form a finite set of ordered pairs D = {(x i , y i ), i = 1, 2,......, n}. Define a grid G of size a × b, divide the number of samples of variables X and Y into a and b parts, calculate the mutual information MI(X, Y) in each cell of the grid G, and take the maximum value of the mutual information calculated in different grid division methods as the mutual information value of the divided grid G. The maximum mutual information of D under the grid G is:

[0122] MI*(D, a, b) = maxMI(D∣G)

[0123] where D∣G represents dividing D using the grid G, and form the feature matrix M(D) with the maximum normalized mutual information values under different divisions a,b

[0124]

[0125] Since different grids G will result in different D∣G, so by performing an exhaustive search on the feature matrix M(D) a,b to obtain the optimal a0×b0 grid G0, the MIC of variables X and Y is defined as:

[0126]

[0127] where B(n) is the maximum grid area for the exhaustive search. It is known from the literature that usually B(n) = n 0.6 has the best effect.

[0128] The value range of MIC is [0, 1]. The larger the value, the stronger the association strength between variables X and Y.

[0129] Cluster analysis

[0130] The variation law of renewable energy power generation is somewhat related to the corresponding weather types. Under different weather conditions, there are also certain differences in its prediction accuracy. Therefore, dividing different weather types according to the wind and light power curves can effectively determine the similar day type to which the period to be predicted belongs, retain the volatility and diversity of historical data within a certain period of time, and thus improve the prediction accuracy. Based on the relatively common K-means clustering algorithm, data clustering analysis is carried out to achieve the division of similar days under different weather types.

[0131] (2) Wind, light, and water renewable resource power generation output prediction model

[0132] The renewable energy power prediction model uses historical power data and factors affecting the change of power values as input parameters, and the power value to be predicted as the output. Then, the power prediction problem can be transformed into a mathematical problem of finding the mapping relationship between the input variable X and the output variable Y. The multi-hidden layer structure of the deep learning network can better extract effective features from a large amount of input data information. It has its unique advantages in dealing with the time series problem of power data itself and the non-linear relationship between various influencing factors and it.

[0133] Wavelet decomposition

[0134] Regarding the strong randomness and volatility of new energy output, and at the same time having a certain periodicity and high time correlation. Wavelet decomposition is used to extract feature data, and the principle of wavelet decomposition is as follows:

[0135] Let any functions g(t) and ψ(t) be square-integrable functions, that is, g(t) ∈ L 2 (R), ψ(t) ∈ L 2 (R), and ψ(t) satisfies the following conditions:

[0136]

[0137] Then the calculation formula of the continuous wavelet transform is:

[0138]

[0139] where ψ p,q (t) is the continuous wavelet generated by the mother wavelet through the scaling variable p and the translation variable q, and t is the time index.

[0140] The reconstruction formula of wavelet transform is:

[0141]

[0142] The discrete wavelet transform (Discrete Wavelet Transform, DWT) is often used to process discrete signals. The discretization formulas for p and q are respectively taken as and The calculation formula of the discrete wavelet is as follows:

[0143]

[0144] where p0 = 2, q0 = 1; The calculation formula of DWT is:

[0145]

[0146] where the input signal passes through the low-pass filter L(e) and the high-pass filter H(e), and further uses the Mallat algorithm, and its calculation formula is as follows:

[0147] a j = a j+1 h1; d j = d j+1 l1 j = 0, 1, L, n - 1

[0148] where h1 and l1 are the low-pass filter coefficient and the high-pass filter coefficient respectively, a j and d j are the low-frequency signal sequence and the high-frequency signal sequence after decomposition respectively, and n is the number of wavelet decomposition layers.

[0149] Its fusion reconstruction formula is:

[0150] a j = a j+1 h2 + d j+1 l2 j = n - 1, L, 1, 0

[0151] where h2 and l2 are the dual operators of h1 and l1 respectively

[0152] Distributed renewable energy power generation prediction framework Figure 3 :

[0153] The specific process of the output prediction model for new types of wind power and photovoltaic power sources is as follows:

[0154] Step1: Preprocess the obtained input data. In order to eliminate the influence of dimensions, eliminate the dimension error, and accelerate the training process, normalize the input training data and convert it into scalar values and output power values.

[0155] Step2: Use the wavelet decomposition method to perform wavelet decomposition and single-branch reconstruction on the input time series data, realize the optimized extraction of trend information, filter out the fluctuation information, separate the noise, and obtain the final input sequence of the model.

[0156] Step 3: Initialize the BiLSTM network based on the Attention mechanism. Input the training data of the decomposed dataset into the network respectively, and use the training samples to train the network. Take the mean squared error as the loss function. The network training adopts the Adam optimization algorithm to update the weights and obtain the prediction model.

[0157] Use the trained neural network model to predict each subsequence decomposed from the time period to be predicted respectively, add the predicted subsequences together, verify the prediction accuracy, and obtain the finally predicted wind power and photovoltaic power data.

[0158] Construct an optimal design model for the converter station's station service power supply combination

[0159] Multi-objective resource screening model

[0160] Decompose the problem into the composition method of coalition members and task allocation. Apply the modern portfolio theory to transform the problem of distributed energy resources (DER) response power allocation into the weights of risk-free and risky asset allocation. By analyzing the influence of DER combination methods on the revenue risk of the converter station, an optimization combination model is established, considering the access cost of coalition members and the incentive and penalty costs of task allocation:

[0161] Converter station resource combination risk assessment

[0162] Due to the heterogeneity of DERs participating in the interaction, the "products" are diverse. The influence of the combination of various types of assets on the expected revenue PER() of the converter station and the diversification effect of revenue risk can be analyzed according to the mean-variance analysis method and the MPT efficient frontier model in the MPT theory:

[0163]

[0164] In the formula, i represents the asset type, w i and E(R i ) respectively represent the investment weight of the i-th type of asset in the portfolio and its expected revenue.

[0165] The revenue risk of the asset portfolio is measured by the variance of the revenue:

[0166]

[0167] In the formula, σ i is the standard deviation of the revenue of the i-th type of asset; ρ ij is the correlation between the revenues of the i-th and j-th types of assets.

[0168] When all the response power provided by the combined users of the converter station is risky assets, the feasible region of the combination can be calculated based on the Markowitz portfolio theory. The feasible region is a set formed by the expected returns - risks under different weights, with the expected return E(R p ) as the vertical axis and the risk - return value σ p as the horizontal axis. For a detailed introduction to the portfolio feasible region, see Appendix A.

[0169] In the feasible region, the returns and risks of each asset itself determine the endpoints of the feasible region; the correlation between resources determines the shape of the feasible region, and the effect of portfolio risk diversification increases as the correlation changes from positive to negative. Inside the feasible region, at various risk levels, the locus formed by connecting the points with the maximum expected return is the efficient frontier of the portfolio. The utility function represents the degree of risk preference of different investors. When the utility is the same and maximum (U is a constant and U = U max ), the indifference curve can be obtained. At the point , the indifference curve is tangent to the portfolio efficient frontier, and the slope of the tangent point According to experience, the value range of A is 2 - 4. For risk - preferring investors, A < 3; for rational investors, A = 3; for risk - averse investors, A > 3.

[0170] When the converter - station portfolio includes risk - free assets, based on the optimal portfolio plan of risky assets, part of the risky assets can be replaced by risk - free assets to reduce the return risk of the portfolio. At this time, the shape of the efficient frontier is a combination of a straight line and a curve, and the straight line is also called the capital allocation line. The straight line passes through the point with an intercept of (0, R f ) representing the expected return of the risk - free asset. According to the two points of the optimal portfolio solution and (0, R f ), another expression of the slope can be obtained, and at this time the slope is called the Sharpe ratio K SR :

[0171]

[0172] To sum up, regarding the converter - station operator as a rational investor, the problem to be solved for the optimal combination of resources is transformed into measuring the returns and risks of different combination methods with the Sharpe ratio as an index among the combination methods that meet the application constraints, and selecting the maximum expected return as the optimal solution. The optimal portfolio plan considering utility maximization has the following characteristics:

[0173]

[0174] Analysis of the output characteristics of the station - use power supply

[0175] According to the control authority of the converter station over resource devices, resources can be divided into two categories: controllable and uncontrollable. In the contract signed between the converter station and users, if the users do not entrust the control authority of the terminal devices to the converter station, they can be regarded as demand response users. The response power provided by such users is uncertain and is regarded as a risky asset. These users are mostly large-load users, such as campuses and office buildings, with various devices that can adjust their electricity consumption characteristics internally and have their own energy management systems. The response actions of such users are similar to incentive-based demand response (IBDR). As an aggregator, the converter station does not need to know the specific response methods of such users, but only needs to focus on the actual response effects of the users. To reasonably optimize the electricity consumption plan of users, such users need to submit a predicted curve of electricity purchase demand to the converter station before the scheduling period. and response range The converter station then believes that the user can achieve any response power within the response capacity. The response characteristics of such resources can be established as:

[0176]

[0177] where, ΔP t i,DR,k , ΔPt is uncertain, which causes revenue risks for the converter station. The response uncertainty of general IBDR users essentially stems from the lack of information and it is difficult to match their own response capabilities with the response requirements of the system; however, under the management mode of the converter station, after comprehensively considering the external demand and the response capabilities of internal resources, the converter station then allocates the response power of each participating user. Therefore, free-response users can achieve global optimality by responding according to the allocation plan of the converter station, and the response deviation will only cause economic losses to the converter station. After considering the uncertainty, ΔP t i,DR,k Rewrite as:

[0178]

[0179] In the formula, is the expected value of the response power allocated to the user, is the random deviation of the user's response power.

[0180] By analyzing the historical response deviation power ε of risky asset users participating in the response, the revenue standard deviation σ of each user can be obtained i and the correlation ρ ij between risky assets, ρ ij > 0 indicates that the response errors of risky asset user i and risky asset user j are consistent and always show over-response or under-response at the same time; conversely, ρ ij < 0 indicates that the response errors between users are complementary.

[0181] Risk-free asset response characteristic model

[0182] In the contract signed between the converter station and the user, if the user entrusts the control authority of the terminal equipment to the converter station, the response power provided by such users is regarded as a risk-free asset. The response power provided by such users is essentially a risk-free asset with σ = 0 and ρ = 0, which can avoid the revenue risk caused by the user's response willingness. Considering that the technical characteristics of the terminal equipment entrusted to the converter station are different, the user-side interactive resources are divided into two categories: flexible controllable load and DG.

[0183] According to the analysis of the energy demand of flexible loads, a response characteristic model of flexible loads is established, requiring that no load shedding events occur during the user response process, that is, ensuring that the reliability indexes R LOCI = 0 and R LOLP = 0. To ensure the reliability of flexible loads, they are divided into shiftable loads, shiftable loads, and interruptible loads according to the dispatching method. The response mechanisms of the three types of flexible loads are analyzed, and the schematic diagrams of the response schemes of the three types of flexible loads are drawn, as shown in Figure 4 shown.

[0184] Reducible load:

[0185] The modeling of reducible load is relatively simple. It only needs to consider whether the load is reduced in each time period and the reduction ratio. We can obtain the following expression:

[0186]

[0187] Among them, the superscript {i,re,k} indicates that the load participating in the response is the k-th reducible load at node i. The 0-1 variable is introduced to indicate whether the reducible load is reduced in the t-th time period, indicating that it is reduced in the t-th time period; α i,re,k , α represents the reduction rate of the reducible load, 0 < α i,re,k < 1; represents the power of the reducible load in the t-th time period before participating in the response of the converter station; P t i,re,k represents the power of the reducible load in the t-th time period after participating in the response of the converter station; kC represents the compensation for the capacity cost of the reducible load; c re represents the reported unit capacity cost of the reducible load; represents the compensation for the electricity price cost of the reducible load; C i,re,k represents the operating cost of the reducible load; T represents the dispatching cycle.

[0188] Considering the actual work and life conditions of users, it is also necessary to impose constraints on the reduction duration of the load that can be curtailed:

[0189] (Ⅰ) Minimum duration constraint:

[0190]

[0191] (Ⅱ) Maximum duration constraint

[0192]

[0193] Among them, and represent the minimum and maximum durations of load reduction respectively; represents the current number of load reduction time periods. The above formulas represent the minimum duration constraint at the end of the previous scheduling period; the minimum duration constraints during and at the end of the scheduling period, the maximum duration constraint of the reduction state at the end of a scheduling period, and the maximum duration of load reduction during the scheduling period.

[0194] Shiftable load

[0195] Shiftable load is a special type of transferable load that only shifts the load curve on the time axis. In this scheduling mode, the power consumption of the load before and after shifting must be the same, ensuring that the reliability index R LOLP = 0. The specific response characteristic model of shiftable load is as follows:

[0196]

[0197] S i,shift,k = [t i,sh-,k ,t i,sh+,k -t i,sh,k +1]∪{t i,sh,k*}

[0198]

[0199] Among them, the superscript {i,shift,k} indicates that the load participating in the response is the k-th shiftable load at node i. represents the response cost of the shiftable load; represents the cost per unit response power of the shiftable load; represents the state of the i-th shiftable load. If indicates that the load starts from time period τ; the interval of shiftable time periods acceptable to the load; is the starting time period of the load; the sum of the power consumption of the shiftable load; represents the total power consumption time period of the shiftable load. The total load includes Si,shift,k Represents the set of start time periods of shiftable loads.

[0200] Since there are two situations for shiftable loads after scheduling, namely not shifting and shifting to an acceptable time period, the following constraint conditions can be constructed:

[0201]

[0202] Shiftable load

[0203] Shiftable loads are relatively flexible. Since there are no restrictions on continuity and timing, different quantities and time intervals can be arranged within the acceptable transfer time interval, so it is difficult to calculate the scheduling compensation cost. Considering that shiftable loads have the highest flexibility, when scheduling shiftable loads, instead of compensating according to the response quantity, users are directly given a total compensation of C i,trans,k However, the load itself needs to be subject to certain constraints:

[0204] (Ⅰ) Minimum duration constraint:

[0205]

[0206] (Ⅱ) Load power constraint:

[0207]

[0208] (Ⅲ) Transferable interval constraint:

[0209]

[0210] (Ⅳ) Power balance constraint:

[0211]

[0212] Where represents the response status of the shiftable load; represents the minimum duration of the i-th type of shiftable load; and represent the upper and lower power limits of the shiftable load respectively; represents the total power demand of the shiftable load before non-response.

[0213] Taking energy storage devices, such as battery energy storage, as typical shiftable loads for research, considering the impact of overcharging and over-discharging on the battery service life, during the participation in the response process, it is required that the state of charge of the battery always remains within a certain range:

[0214]

[0215] In the formula, Indicates the state of charge of the ESS at time t. and Indicate the upper and lower limits of the state of charge allowed for the ESS user within the scheduling period, respectively set

[0216] Distributed power source

[0217] DG users include distributed photovoltaics, distributed wind power generation, small hydropower, biomass power generation, etc. Among them, distributed photovoltaics and wind power generation are restricted by meteorological conditions. For example, the DC side of a photovoltaic grid-connected inverter usually adopts the MPPT strategy to control the active power output to the grid connection point. The MPPT power represents the maximum output of the DG. Under the maximum power limit, the response characteristics of the DG are similar to those of the curtailable load, and the active output can be reduced below the MPPT power. However, from the perspective of the power grid, the reduction in the DG output is equivalent to an increase in the power purchase volume of the converter station, which is regarded as providing positive response power. The response characteristics of the DG are:

[0218]

[0219] Among them, the superscript {i,DG,k} indicates that the load participating in the response is the k-th DG at node i. Introduce a 0-1 variable Indicates whether the DG is reduced, Indicates being reduced in the t time period; β i,DG,k Indicates the reduction rate of the curtailable load; Represents the response cost of the DG; Indicates the DG capacity cost compensation; Indicates the curtailable load electricity price cost compensation; Represents the cost per unit response power of the DG;

[0220] Converter station resource optimization combination model

[0221] The resource optimization combination problem of the converter station is similar to the unit commitment problem. However, the unit commitment problem focuses on the power generation capacity of the units, while the converter station focuses on the response capacity of the resources. An inappropriate converter station resource combination method cannot meet the requirements of the application for its product characteristics and is also difficult to obtain good economic efficiency.

[0222] Taking the converter station participating in the day-ahead energy market as an example, an optimization combination model is established:

[0223] max:E(R p )=C in -E(C cost )

[0224]

[0225] In the model, with the goal of maximizing the expected revenue, in the objective function, E(R p ) represents the expected revenue of the combined converter station; C in represents the expected revenue of the converter station. After the market clearing determines the bidding curve P aim,t , the expected revenue C in from the energy market is C r - C w )P aim,t . Δt is a determined constant value. The profit margin comes from the difference between the retail electricity price C r and the wholesale electricity price C w in the power market. E(C cost ) is the expected cost, which consists of three parts, and the expression is as follows:

[0226]

[0227] In the formula, f1 is the incentive cost required to encourage users to change their output plans and provide responsive electricity; f2 is the uncertainty cost caused by the deviation between the actual power of the converter station and the target power, corresponding to the penalty cost from the power market during the settlement stage, and c er is the unit deviation penalty price; f3 is the access cost for communication and computing resources required to maintain users' participation in the interaction of the converter station. C i,j,k is the unit response incentive cost of the interactive resource, that is, the internal trading electricity price; λ is the access cost required for a single resource to participate in regulation; V i,j,k is a 0 / 1 variable, and V i,j,k = 1 indicates that this resource participates in the response. In the constraint conditions, h(P t i,j,k ) ≥ 0 represents the risk and utility constraint in the optimal combination scheme, u(P t i,j,k ) ≥ 0 represents the resource response characteristic constraint shown in the formula), g(P t i,j,k ) = 0 is the equality constraint representing the aggregated external characteristics of the converter station, and s(P t i,j,k ) ≥ 0 represents the application constraint. And according to the combined expected revenue expression in formula (4 - 36), the weights of various asset combinations in the formula can be expressed as

[0228] Analysis of Mixed Decision Variables

[0229] For the solution of LSMOPs using the divide-and-conquer method, classifying and optimizing decision variables separately is a commonly used method. However, for different problems, simply relying on fixed classification criteria sometimes cannot completely separate convergence variables and diversity variables. Especially for some decision variables that can control both convergence and diversity, it is difficult to classify them completely. Therefore, based on the large-scale multi-objective optimization algorithm using an adaptive localized control variable analysis approach (LSMOEA / D), a hybrid decision variable analysis method is proposed and improved. Instead of qualitatively classifying decision variables as convergence variables or diversity variables, the control property ranking CPS of each decision variable is obtained under the dual mechanisms of angle and distance. The smaller the CPS value, the more the decision variable tends to control the convergence of the population, and the larger the serial number, the more it tends to control the diversity of the population.

[0230] To make the CPS represent the control properties of convergence and diversity, the CPS formula uses two factors, angle and distance, to act together, so as to improve the accuracy of its control properties. Taking Figure 7 as an example, to calculate the CPS value of the decision variable x i in a two-objective optimization problem, C i is a weight center after the clustering operation of all weight vectors. First, two random candidate solutions are selected from the neighborhood of C i , then the selected candidate solutions are perturbed 5 times, and the perturbed solutions of each group are normalized to obtain two lines L1 and L2. Two CPS values are obtained using the CPS formula, and finally the average value is taken to obtain the CPS value of the decision variable x i . The CPS formula is as follows:

[0231]

[0232] Specifically, the angle θ refers to the angle between the normalized straight line of the perturbed solution and the current weight center vector Ci, and the distance d refers to the distance between the two points with the farthest distance among the perturbed solutions in the direction of the weight center vector. The smaller the angle, the more the decision variable tends to control the convergence of the population, and the larger the angle, the more the decision variable tends to control the diversity of the population; the smaller the distance, the greater the amplitude of each perturbation of the decision variable and the longer the step length of the population change, so it is more beneficial to search for both the convergence and diversity of the population. The combined action of the two search factors can effectively improve the convergence speed and distribution of the optimized population. As the core content of the analysis of the mixed decision variable, the pseudocode of the control attribute sorting is as shown in the algorithm. First, cluster all the weight vectors to obtain k weight centers C, then select nSel solutions from the neighborhood of each weight center for nPer perturbations to obtain the solution set S, fit S to generate the line L, calculate the value of CPS using the formula, then calculate the mean of the generated nSel CPS values, and finally sort the results to obtain CPSIndex.

[0233] Algorithm flow

[0234] First, randomly initialize the population P and the reference point z * , generate uniform weight vectors, then obtain CPSIndex. In the population evolution stage, randomly select parents for each weight vector, use the genetic algorithm to generate offspring individuals, then update the reference point and the neighborhood, and use the boundary crossover aggregation method based on the penalty factor to update the population until the algorithm termination condition is met.

[0235] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0236] It should be understood that the detailed description of the technical solutions of the present invention with the aid of the preferred embodiments above is illustrative rather than restrictive. Those of ordinary skill in the art can modify the technical solutions described in each embodiment based on reading the specification of the present invention, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. An optimization model and method for the station service power supply of a converter station, characterized in that: The method described above includes: Construct an energy consumption characteristic analysis and power consumption prediction model for converter stations, and adopt an energy consumption characteristic analysis of user converter stations based on wavelet transform and optimized fast density peak clustering algorithm, as well as a power consumption prediction model for converter stations based on the XGBoost algorithm; Construct a prediction model for the output of new power sources such as wind power and photovoltaic power, and adopt a wavelet decomposition - bidirectional long short - term memory network based on the Attention mechanism as the prediction model; Combine the power consumption prediction model and the new power source output prediction model to construct an optimized design model for the converter station's station - used power supply combination.

2. The optimized model and method for the station service power supply of a converter station according to claim 1, wherein: The construction of the energy consumption characteristic analysis and power consumption prediction model for converter stations includes: Power consumption prediction for converter stations based on the XGBoost algorithm; Energy consumption characteristic analysis of user converter stations based on wavelet transform and optimized fast density peak clustering algorithm. Design an optimized fast density peak clustering algorithm based on wavelet transform, which is an improved algorithm by combining wavelet transform, KNN algorithm, fast density peak clustering algorithm, and clustering center automatic selection strategy.

3. The optimized model and method for the station service power supply of a converter station according to claim 2, wherein: The energy consumption characteristic analysis of user converter stations by the wavelet transform and optimized fast density peak clustering algorithm respectively includes wavelet transform, clustering analysis, and reconstruction analysis; For the wavelet transform, the sample data is subjected to wavelet transform and decomposed into spectral components with different time scales. The spectral components obtained with different decomposition levels are different. Each time of decomposition reduces the time scale by half. After decomposing n layers, the detail components D, D…, D, and the approximation component A can be obtained; For the clustering analysis, the spectral components D, D…, D, and A after wavelet decomposition are used as inputs for clustering analysis respectively. Find the K nearest neighbors of each sample point, calculate the local density p and KNN distance δ of each sample point using the optimized fast density peak clustering algorithm, so as to obtain the decision variable, and then use the clustering center automatic selection strategy to obtain the clustering centers of different spectral components. Finally, mark the category to which all sample points belong; For the reconstruction analysis, reconstruct the clustering centers of each spectral component. The clustering centers of different spectral components are denoted as TLP and TLP, and the typical load characteristic curve of the original data is TLP. By analyzing the reconstructed typical load curve, the electricity consumption characteristics of users can be understood.

4. The optimized model and method for the station service power supply of a converter station according to claim 1, characterized in that: The construction of the prediction model for the output of new power sources such as wind power and photovoltaic power includes: Cleaning of wind, light, and water monitoring data, including: abnormal data detection, autocorrelation and partial autocorrelation analysis, and correlation analysis between power generation power and meteorological factors; For the power output prediction model of wind, light, and water renewable resources, the renewable energy power prediction model uses historical power data and factors affecting the change of power values as input parameters, and the power value to be predicted as the output. Then the power prediction problem is transformed into a mathematical problem of finding the mapping relationship between the input variable X and the output variable Y. Adopt a wavelet decomposition - bidirectional long short - term memory network based on the Attention mechanism as the prediction model.

5. The optimized model and method for the station service power supply of a converter station according to claim 1, characterized in that: The optimized design model for the converter station's station - used power supply combination includes: A multi - objective resource screening model. Decompose the problem into the composition method of coalition members and task allocation, and use modern portfolio theory to transform the response power allocation problem into the allocation weights of risky and risk - free assets; The risk assessment of the converter station resource combination is analyzed according to the mean-variance analysis method and the MPT efficient frontier model in the MPT theory; Analysis of the output characteristics of the station service power supply: According to the control authority of the converter station over the resource equipment, the resources are divided into two categories: controllable and uncontrollable; in the contract signed between the converter station and the user, if the user does not entrust the control authority of the terminal equipment to the converter station, it is regarded as a demand response user. The response power provided by such users is uncertain and is regarded as a risky asset. Such users are mostly large load users, with various adjustable power consumption characteristic devices inside and their own energy management systems; Response characteristic model of risk-free assets: In the contract signed between the converter station and the user, if the user entrusts the control authority of the terminal equipment to the converter station, the response power provided by such users is regarded as a risk-free asset; the response power provided by such users is essentially a risk-free asset with σ = 0 and ρ = 0, avoiding the income risk caused by the user's response willingness. Considering that the technical characteristics of the terminal equipment entrusted to the converter station are different, the user-side interactive resources are divided into two categories: flexible controllable load and DG.

6. The optimized model and method for the station service power supply of a converter station according to claim 5, characterized in that: The above-mentioned optimal design model for the converter station's station service power supply combination also includes: Optimal combination model of converter station resources: The following optimal combination model is established: max:E(R p ) = C in -E(C cost ) In the model, with the goal of maximizing the expected revenue, in the objective function, E(R p ) represents the expected revenue of the combined converter station; C in represents the expected revenue of the converter station. After the market clearing determines the bidding curve P aim,t , the expected revenue C in = (C r - C w )P aim,t Δt is a determined constant value. The profit margin comes from the difference between the retail electricity price C r and the wholesale electricity price C w in the electricity market; E(C cost ) is the expected cost, which consists of three parts and the expression is as follows: where f1 is the incentive cost required to motivate users to change the output plan and provide responsive power; f2 is the uncertainty cost caused by the deviation between the actual power and the target power of the converter station, corresponding to the penalty cost from the power market during the settlement stage, and c er is the unit deviation penalty price; f3 is the access cost to communication and computing resources required to maintain user participation in the interaction of the converter station; C i,j,k is the unit response incentive cost of the interactive resource, that is, the internal trading electricity price; λ is the access cost required for a single resource to participate in regulation; V i,j,k is a 0 / 1 variable, and V i,j,k = 1 indicates that the resource participates in the response; in the constraint conditions, h(P t i,j,k ) ≥ 0 represents the risk and utility constraints in the optimal combination plan, and u(P t i,j,k ) ≥ 0 represents the resource response characteristic constraints shown in the formula), and g(P t i,j,k ) = 0 is the equality constraint representing the aggregated external characteristics of the converter station, and s(P t i,j,k ) ≥ 0 represents the application constraint. The weights of various asset combinations in the formula are expressed as 7. The optimized model and method for the station service power supply of a converter station according to claim 5, characterized in that: The solution of the above-mentioned optimal design model for the converter station's station service power supply combination adopts an optimization heuristic solution algorithm.

8. The optimized model and method for the station service power supply of a converter station according to claim 5, characterized in that: The specific process of the output power prediction model of the new wind power and photovoltaic power sources is as follows: Step 1: Preprocess the obtained input data. In order to eliminate the influence of dimensions, eliminate the dimension error, and accelerate the training process, the input training data is normalized and converted into scalar values and output power values; Step 2: Use the wavelet decomposition method to perform wavelet decomposition and single-branch reconstruction on the input time series data, realize the optimal extraction of trend information, filter the fluctuation information, separate the noise, and obtain the final input sequence of the model; Step 3: Initialize the BiLSTM network based on the Attention mechanism, input the decomposed dataset training data into the network respectively, and use the training samples to train the network; use the mean square error as the loss function, and the network training adopts the Adam optimization algorithm to update the weights to obtain the prediction model; Use the trained neural network model to predict each subsequence decomposed during the prediction period respectively, add the predicted subsequences, verify the prediction accuracy, and obtain the final predicted wind power and photovoltaic power data.

Images