XGBoost-based demand response potential dynamic evaluation method

Abstract

The invention discloses an electric vehicle demand response potential dynamic evaluation method. The method specifically comprises the following steps: step 1, acquiring historical charging load data of an electric vehicle in a region, and preprocessing the historical load data; step 2, analyzing key influence factors of electric vehicle demand response potential and a mechanism of participating in demand response; 3, constructing a demand response dynamic evaluation model through the key influence factors and the demand response mechanism, and carrying out training and hyper-parameter optimization on the demand response dynamic evaluation model based on historical load data; and step 4, obtaining an electric vehicle demand response potential value through the demand response dynamic evaluation model trained in the step 3. According to the method, the demand response potential of the electric vehicle can be dynamically evaluated in advance.

Description

Technical Field

[0001] The present invention belongs to the technical field of demand response potential evaluation, and specifically relates to a demand response potential dynamic evaluation method based on XGBoost. Background Art

[0002] Large-scale electric vehicle access provides more flexible resources for the distribution network. Through in-depth research on electric vehicles' participation in grid demand response control technology, electric vehicles can actively participate in the peak load regulation of the distribution network, realize orderly charging, and help steadily advance the dual carbon goals. Electric vehicle charging loads are spatiotemporally flexible and mobile. By setting reasonable electricity prices and guiding the charging behavior of electric vehicle users, electric vehicle users can actively participate in the grid demand response. As an important part of power demand response, electric vehicle demand response has been widely used in emergency power supply, new energy consumption, peak load regulation, and power flow optimization.

[0003] The demand response behavior of electric vehicles is quite different from the traditional demand response behavior. The travel of electric vehicles is random and flexible, and their charging demand is separated from their travel demand in time, that is, adjusting their charging time during the vehicle's downtime will not affect their travel behavior. Moreover, most electric vehicles are down most of the time, so their demand response potential is huge. The coordination between the charging demand and response behavior of electric vehicles and grid dispatch is a challenge for the development of new power systems. How to utilize the flexibility of electric vehicles and make them an important participant in power system dispatch has become an urgent problem to be solved. Conventional demand response potential assessment methods are mostly based on probabilistic statistical methods, assuming that users meet specific probability distributions. If the assumption is unreasonable, the assessment results may have large deviations. The demand response potential obtained based on traditional methods is static and ex post facto, and the accuracy is not high. Therefore, there is a lack of a comprehensive consideration of multiple variable factors, which can be used to conduct a pre-, dynamic and accurate assessment of the demand response potential of electric vehicles, so as to better play the mobile energy storage characteristics of electric vehicles. Summary of the invention

[0004] The purpose of the present invention is to provide a dynamic evaluation method for demand response potential based on XGBoost, which can perform ex ante and dynamic evaluation on the demand response potential of electric vehicles.

[0005] The technical solution adopted by the present invention is a dynamic evaluation method of demand response potential based on XGBoost, which specifically includes the following steps: Step 1, obtaining historical charging load data of electric vehicles in the region and preprocessing the historical load data; Step 2: Analyze the key factors affecting the demand response potential of electric vehicles and the mechanisms involved in demand response; Step 3: construct a demand response dynamic evaluation model through key influencing factors and demand response mechanism, and train and optimize hyperparameters of the demand response dynamic evaluation model based on historical load data; Step 4, obtaining the electric vehicle demand response potential value by completing the demand response dynamic evaluation model trained in step 3.

[0006] The present invention is also characterized in that The preprocessing of historical load data in step 1 includes cleaning, filling in gaps, standardization and division of data; Cleaning uses outlier processing and redundant data deduplication; Specifically, when the missing data of load data accounts for more than 20%, it is considered that the data has caused the loss of valid information and the data is directly deleted. The missing data of the remaining load data are supplemented based on the historical data of other dates and time periods. Standardization is to normalize the data, as shown in the following formula: ; The historical load data is divided into training set, validation set and test set.

[0007] To supplement missing values, first calculate the mean of the historical data of other dates in this period to obtain the historical mean, as shown in the following formula: ; In the formula, For the front The average value of the missing data points in each week; For the The value of the week corresponding to the point where data is missing; Then calculate the average value of the charging station load at adjacent times to obtain the average value of the adjacent points, as shown in the following formula: ; In the formula, The data before and after the missing point The mean value of load data at each time point; , Before and after the data missing point Load data at the moment; Based on the historical mean and the mean of neighboring points, the filling value of the missing data point is calculated by weighted estimation method. The specific formula is as follows: ; In the formula, is the historical load mean of the data missing point; is the mean load of the neighboring point where data is missing; is the distribution weight of the historical load mean and the neighboring load mean, .

[0008] The key factors affecting the demand response potential of electric vehicles in step 2 include the starting charging time, starting SOC, charging time, charging power, SOC at the end of charging, electricity price policy, energy policy, and charging infrastructure construction; The mechanism of electric vehicle demand response potential participating in demand response is that charging stations set electricity prices, and users choose whether to participate in the response based on the electricity prices set by charging stations. If users choose to participate in the response, their charging behavior will be adjusted.

[0009] In the mechanism of participating in demand response, users choose whether to participate in the response based on the electricity price set by the charging station. There are price-based demand response and incentive-based demand response. The price includes real-time electricity price and time-of-use electricity price, and the incentive includes user subsidies given by the charging station.

[0010] The objective function of the demand response dynamic evaluation model in step 3 is as follows: ; In the formula, and Represents the number of leaf nodes and the regularization penalty coefficient; T represents the number of leaf nodes in the tree.

[0011] Training the demand response dynamic evaluation model specifically includes the following steps: S301, label each load data in the training set. The load data on the demand response day is , the load on non-demand response days is , the labeled load data is input into the input layer of the demand response dynamic evaluation model as a training sample set, as shown in the following formula: ; In the formula, , is the sample weight; At the same time, expressions of key influencing factors and demand response mechanisms are constructed, as shown in the following formula, and the expressions are input into the demand response dynamic evaluation model; ; In the formula, is the demand response factor set expression, is the weight of the influencing factor on the training data sample; S302, constructing the first decision tree with the key influencing factors as nodes, initializing a random node, calculating the separation threshold of the node, and updating the value of the node based on the separation threshold to minimize the loss function; S303, select the key influencing factors and split points that minimize the loss function as the splitting conditions of the current node to split and generate new child nodes, and calculate the feature gain of the key influencing factors, as shown in the following formula: ; S304, repeat S303 for the newly generated child nodes until the depth of the decision tree meets the requirements, at which time the sum of the node scores of the decision tree is used as the sample prediction value, and the residual is calculated based on the sample prediction value; S305, based on the characteristic gain calculated in S303, the weights of the key influencing factors of the demand response are obtained, and a new decision tree is constructed to fit the residual of the previous prediction by taking the key influencing factor set of the node and the corresponding residual as input; S306, combining the predicted value of the new decision tree with the predicted value of the current demand response dynamic evaluation model to complete the update of the demand response dynamic evaluation model until the training is completed. tree; S307, repeat S303 to S306 until the prediction effect of the demand response dynamic evaluation model meets the set standard or reaches the set maximum number of iterations.

[0012] The hyperparameter optimization of the demand response dynamic evaluation model is to use the sklearn machine learning algorithm package to perform grid search and As an evaluation indicator, it is shown in the following formula: .

[0013] The validation set is input into the demand response dynamic evaluation model that has completed training and hyperparameter optimization for verification. If the verification fails, retraining is performed.

[0014] Step 4 specifically inputs the test set into the verified demand response dynamic evaluation model to obtain the load value after the demand response, and obtains the demand response potential value based on the obtained load value minus the baseline load value.

[0015] The beneficial effects of the present invention are: The XGBoost-based demand response potential dynamic evaluation method of the present invention analyzes the influencing factors of demand response potential from different angles, can realize dynamic evaluation of the response potential of electric vehicles in different functional areas, at different times and on different dates, and can obtain the demand response potential within the next 24 hours in advance and dynamically, with higher accuracy and reliability. The evaluation results provide a strong reference basis for subsequent scheduling strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the dynamic evaluation method of demand response potential based on XGBoost of the present invention. Figure 2 This is an analysis diagram of the XGBoost mechanism in the present invention. DETAILED DESCRIPTION

[0017] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The present invention provides a dynamic evaluation method for demand response potential based on XGBoost, such as Figure 1 As shown, the specific steps include: Step 1, obtaining historical charging load data of electric vehicles in the region and preprocessing the historical load data; Furthermore, the preprocessing of historical load data in step 1 includes cleaning, filling in gaps, standardizing and dividing the data; Cleaning uses outlier processing and redundant data deduplication; Specifically, when the missing data of load data accounts for more than 20%, it is considered that the data has caused the loss of valid information and the data is directly deleted. The missing data of the remaining load data are supplemented based on the historical data of other dates and time periods. Standardization is to normalize the data, as shown in the following formula (1): (1); The historical load data is divided into training set, validation set and test set.

[0019] Furthermore, the missing value supplementation first calculates the mean of the historical data of other dates in this period to obtain the historical mean, as shown in the following formula (2): (2); In the formula, For the front The average value of the missing data points in each week; For the The value of the week corresponding to the point where data is missing; Then the average value of the charging station load at adjacent times is calculated to obtain the average value of the adjacent points, as shown in the following formula (3): (3); In the formula, The data before and after the missing point The mean value of load data at each time point; , Before and after the data missing point Load data at the moment; Based on the historical mean and the mean of neighboring points, the filling value of the missing data point is calculated by weighted estimation method. The specific formula is as follows (4): (4); In the formula, is the historical load mean of the data missing point; is the mean load of the neighboring point where data is missing; is the distribution weight of the historical load mean and the neighboring load mean, .

[0020] Step 2: Analyze the key factors affecting the demand response potential of electric vehicles and the mechanisms involved in demand response; Step 3: construct a demand response dynamic evaluation model through key influencing factors and demand response mechanism, and train and optimize hyperparameters of the demand response dynamic evaluation model based on historical load data; Step 4, obtaining the electric vehicle demand response potential value by completing the demand response dynamic evaluation model trained in step 3.

[0021] Furthermore, the test set is specifically input into a verified demand response dynamic evaluation model to obtain a load value after demand response, and a demand response potential value is obtained based on the obtained load value minus the baseline load value.

[0022] The baseline load value is not protected by the present invention, but belongs to the existing mature prior art, and the specific reference is made to the solution with the publication number of CN117933468A. The calculation model of the baseline load value is located at the output layer when constructing the demand response dynamic evaluation model.

[0023] Example 2 On the basis of Example 1, the key influencing factors of the electric vehicle demand response potential in step 2 of the demand response potential dynamic evaluation method based on XGBoost of the present invention include two aspects: subjective perspective and objective perspective.

[0024] From a subjective perspective, when, where and how to charge is determined by the owner's behavior and driving habits. The demand response potential of electric vehicles is related to their starting charging time, starting SOC, charging time, charging power and SOC requirements at the end of charging.

[0025] From an objective perspective, regional electricity price policies, energy policies, and charging infrastructure construction will have an impact on user decisions, which in turn will affect the demand response potential of electric vehicles.

[0026] The mechanism of electric vehicle demand response potential participation in demand response refers to the process in which electric vehicles adjust their charging behavior according to grid demand in grid load management to balance grid load and optimize energy utilization. The specific steps are as follows: S201, first the charging station receives relevant policies from the dispatching and control center, formulates relevant peak-shaving electricity prices and compensation measures, and attracts users to participate in demand response.

[0027] S202, electric vehicle users then decide whether to participate in demand response based on the peak-shaving electricity price and subsidy measures of each type of charging station, according to their personal needs and cost benefits.

[0028] S203, if the user decides to participate in demand response, the user's charging behavior will be adjusted accordingly, such as delaying the start time of charging, shortening the charging time, reducing the charging power, etc., so as to avoid the peak load period of the power grid, reduce the pressure on the power grid, and improve energy efficiency.

[0029] After the load participates in demand response, the user's charging behavior and load conditions will be monitored and fed back in real time. These data can be used to estimate the load of electric vehicles after participating in demand response and compare it with the baseline load before participating in demand response to evaluate the demand response potential of electric vehicles.

[0030] S202 is mainly divided into two categories: price-based demand response (PBDR) and incentive-based demand response (IBDR). PBDR mainly includes real-time electricity prices and time-of-use electricity prices, and mainly maintains the balance of supply and demand by flexibly guiding users' charging and discharging behaviors; IBDR attracts users to participate in demand response by giving them certain subsidies or signing contracts with them.

[0031] Furthermore, whether the user chooses to participate in the response based on the electricity price set by the charging station needs to be judged based on the demand response analysis model, which includes a demand response model based on price improvement elasticity theory and a demand response model based on statistical theory.

[0032] The demand response model based on price improvement elasticity theory is established as follows.

[0033] The specific elasticity coefficient of electricity demand price is shown in the following formula (5): (5); in, is the price elasticity coefficient of electricity demand; For demand; For initial demand; For electricity prices; is the initial electricity price.

[0034] Consumers' sensitivity to price changes can be expressed in matrix form, as shown in the following formula (6): (6); in, for Dimensional demand change vector; for dimensional elasticity matrix; for dimensional price change vector, indicating The change of electricity price in a price cycle.

[0035] The demand change factors are ,in For the moment The initial demand, For the moment The corresponding modified demand after consumers respond to price signals; the price change factor is ,in For the moment The initial electricity price, For the moment New electricity prices.

[0036] The price elasticity matrix can be expressed as follows (7): (7); Among them, the matrix diagonal terms is the self-elastic coefficient; the non-diagonal terms of the matrix is the cross elastic coefficient. It means the time When a price change occurs, users Their needs change all the time and are always negative; cross elasticity coefficient Refers to the moment When a price change occurs, users Change their requirements and always take positive values.

[0037] According to the load characteristics, users usually respond to changes in electricity prices to adjust their electricity consumption in the current period. However, users are not very responsive to load shifts caused by changes in electricity prices in other time periods. Therefore, the impact of electricity price changes in the current time period on electricity consumption changes is greater than that in other time periods. Therefore, in order to reflect the impact of prices in different time periods on electricity consumption, the elasticity coefficient weight is introduced, as shown in the following formula (8): (8); in, is the elastic coefficient weight; For the period The period caused by the change of electricity price The change of electric quantity; For the period The total power change is determined by the following equations (9) and (10): (9); (10); By using formula (8) to modify (5) to (7), we can obtain the improved price elasticity matrix as follows: (11); (12); Therefore, use the time Dynamic electricity prices, considering their impact on other periods The impact of can be calculated at the moment The power demand response is as shown in the following formula (13): (13); The demand response model based on statistical theory is mainly used for large industrial users, who have a large load reduction and are more willing to participate in demand response. Data is generally collected through electricity meters or questionnaires.

[0038] Example 3 Based on Example 1, in the dynamic evaluation method of demand response potential based on XGBoost of the present invention, step 3 is specifically as follows: Figure 2 As shown in the figure, the gradient enhanced decision tree algorithm (EXtreme Gradient Boosting, XGBoost) uses the gradient descent method to generate a new tree based on the original tree and make the objective function as small as possible.

[0039] Specifically, the objective function expression of XGBoost is as follows (14) and (15): (14); in, For the The predicted value of samples; For the The actual value of samples; is the number of trees; For the A model of a tree.

[0040] (15); in, and Represents the number of leaf nodes and the regularization penalty coefficient; T represents the number of leaf nodes in the tree; Represents the score of the leaf node.

[0041] Therefore, the objective function of XGBoost can be rewritten as follows (16):

[0042] (16); in, is a constant term.

[0043] XGBoost continuously searches for To optimize the objective function, the objective function of performing second-order Taylor expansion and removing high-order infinitesimal terms is as follows (17) and (18): (17); (18); in, , , then the objective function is as follows (19): (19); Taking the derivative of the objective function, we get Substituting it into equation (19), the final objective function is as follows (20): (20); Furthermore, training the demand response dynamic evaluation model specifically includes the following steps: S301, label each load data in the training set. The load data on the demand response day is , the load on non-demand response days is , the labeled load data is input into the input layer of the demand response dynamic evaluation model as a training sample set, as shown in the following formula (21): (twenty one); In the formula, , is the sample weight; At the same time, expressions of key influencing factors and demand response mechanisms are constructed, as shown in the following formula (22), and the expressions are input into the demand response dynamic evaluation model; (twenty two); In the formula, is the demand response factor set expression, is the weight of the influencing factor on the training data sample.

[0044] Therefore, the input layer is I={ | }, any for( , )-dimensional data, where is the number of training data samples, is the time series length of the input data.

[0045] S302, constructing the first decision tree with the key influencing factors as nodes, initializing a random node, calculating the separation threshold of the node, and updating the value of the node based on the separation threshold to minimize the loss function; S303, select the key influencing factor and the split point that minimizes the loss function as the splitting condition of the current node to split and generate a new child node, and calculate the feature gain of the key influencing factor, as shown in the following formula (23): (twenty three); In the formula, is the characteristic gain of the influencing factor; and Represents the number of leaf nodes and the regularization penalty coefficient; T represents the number of leaf nodes in the tree; For Node The cumulative sum of the first-order partial derivatives of the included samples; For Node The accumulated sum of the second-order partial derivatives of the included samples.

[0046] S304, repeat S303 for the newly generated child nodes until the depth of the decision tree meets the requirements. At this time, the sum of the node scores of the decision tree is used as the sample prediction value, and the residual is calculated based on the sample prediction value; the residual is specifically the difference between the sample prediction value and the true value in the historical data.

[0047] S305, based on the characteristic gain calculated in S303, the weights of the key influencing factors of the demand response are obtained, and a new decision tree is constructed to fit the residual of the previous prediction by taking the key influencing factor set of the node and the corresponding residual as input; S306, combining the predicted value of the new decision tree with the predicted value of the current demand response dynamic evaluation model to complete the update of the demand response dynamic evaluation model until the training is completed. tree; S307, repeat S303 to S306 until the prediction effect of the demand response dynamic evaluation model meets the set standard, that is, the value of the objective function reaches the minimum value, or reaches the set maximum number of iterations.

[0048] Furthermore, the hyperparameters of the demand response dynamic evaluation model are optimized by using the sklearn machine learning algorithm package to perform a grid search method and As an evaluation index, it is shown in the following formula (24): (twenty four).

[0049] In the formula, For the The actual value of samples; for The predicted value of the tree model; For the The average of the tree models.

[0050] The optimal parameters of the demand response dynamic evaluation model in this embodiment are shown in the table:

[0051] n_estimators and max_depth are the two most critical hyperparameters. When they take the default values, adjust other parameters. The value of fluctuates in the range of 50% to 70%. After adjustment, other parameters use the default values, and the optimal The value is 83% ( The closer it is to 100%, the better the model training result is).

[0052] Finally, the validation set is input into the demand response dynamic evaluation model that has completed training and hyperparameter optimization for verification. The value fluctuates between 50% and 70%. If the verification fails, retraining is performed.

Claims

1. A dynamic evaluation method for electric vehicle demand response potential, characterized in that: The specific steps include: Step 1, obtaining historical charging load data of electric vehicles in the region and preprocessing the historical load data; Step 2: Analyze the key factors affecting the demand response potential of electric vehicles and the mechanisms involved in demand response; Step 3: construct a demand response dynamic evaluation model through key influencing factors and demand response mechanism, and train and optimize hyperparameters of the demand response dynamic evaluation model based on historical load data; Step 4, obtaining the electric vehicle demand response potential value by completing the demand response dynamic evaluation model trained in step 3.

2. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The preprocessing of the historical load data in step 1 includes cleaning, filling in gaps, standardizing and dividing the data; The cleaning adopts outlier processing and redundant data deduplication; Specifically, when the missing data of the load data accounts for more than 20%, it is determined that the data has caused the loss of valid information, and the data is directly deleted. The missing data of the remaining load data are supplemented based on the historical data of the same period on other dates; The standardization is to normalize the data, as shown in the following formula: ； The division is to divide the historical load data into a training set, a validation set and a test set.

3. The method for dynamic evaluation of electric vehicle demand response potential according to claim 2, characterized in that: The missing value supplementation first calculates the mean of the historical data of other dates and time periods to obtain the historical mean, as shown in the following formula: ； In the formula, For the front The average value of the missing data points in each week; For the The value of the week corresponding to the point where data is missing; Then calculate the average value of the charging station load at adjacent times to obtain the average value of the adjacent points, as shown in the following formula: ； In the formula, The data before and after the missing point The mean value of load data at each time point; , Before and after the data missing point Load data at the moment; Based on the historical mean and the mean of neighboring points, the filling value of the missing data point is calculated by weighted estimation method. The specific formula is as follows: ； In the formula, is the historical load mean of the data missing point; is the mean load of the neighboring point where data is missing; is the distribution weight of the historical load mean and the neighboring load mean, .

4. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The key factors affecting the demand response potential of electric vehicles in step 2 include the starting charging time, the starting SOC, the charging time, the charging power, the SOC at the end of charging, the electricity price policy, the energy policy, and the construction of charging infrastructure; The mechanism of electric vehicle demand response potential participating in demand response is that charging stations set electricity prices, and users choose whether to participate in the response based on the electricity prices set by charging stations. If users choose to participate in the response, their charging behavior will be adjusted.

5. The method for dynamic evaluation of electric vehicle demand response potential according to claim 4, characterized in that: In the mechanism of participating in demand response, users choose whether to participate in the response based on the electricity price set by the charging station. There are price-based demand response and incentive-based demand response. The price includes real-time electricity price and time-of-use electricity price, and the incentive includes user subsidies given by the charging station.

6. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The objective function of the demand response dynamic evaluation model in step 3 is as follows: ； In the formula, and Represents the number of leaf nodes and the regularization penalty coefficient; T represents the number of leaf nodes in the tree.

7. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: Training the demand response dynamic evaluation model specifically includes the following steps: S301, label each load data in the training set. The load data on the demand response day is , the load on non-demand response days is , the labeled load data is input into the input layer of the demand response dynamic evaluation model as a training sample set, as shown in the following formula: ； In the formula, , is the sample weight; At the same time, expressions of key influencing factors and demand response mechanisms are constructed, as shown in the following formula, and the expressions are input into the demand response dynamic evaluation model; ； In the formula, is the demand response factor set expression, is the weight of the influencing factor on the training data sample; S302, constructing the first decision tree with the key influencing factors as nodes, initializing a random node, calculating the separation threshold of the node, and updating the value of the node based on the separation threshold to minimize the loss function; S303, select the key influencing factors and split points that minimize the loss function as the splitting conditions of the current node to split and generate new child nodes, and calculate the feature gain of the key influencing factors, as shown in the following formula: ； S304, repeat S303 for the newly generated child nodes until the depth of the decision tree meets the requirements, at which time the sum of the node scores of the decision tree is used as the sample prediction value, and the residual is calculated based on the sample prediction value; S305, based on the characteristic gain calculated in S303, the weights of the key influencing factors of the demand response are obtained, and a new decision tree is constructed to fit the residual of the previous prediction by taking the key influencing factor set of the node and the corresponding residual as input; S306, combining the predicted value of the new decision tree with the predicted value of the current demand response dynamic evaluation model to complete the update of the demand response dynamic evaluation model until the training is completed. tree; S307, repeat S303 to S306 until the prediction effect of the demand response dynamic evaluation model meets the set standard or reaches the set maximum number of iterations.

8. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The hyperparameter optimization of the demand response dynamic evaluation model is to use the sklearn machine learning algorithm package to perform grid search and As an evaluation indicator, it is shown in the following formula: 。 9. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The validation set is input into the demand response dynamic evaluation model that has completed training and hyperparameter optimization for verification. If the verification fails, retraining is performed.

10. The method for dynamic evaluation of electric vehicle demand response potential according to claim 1, characterized in that: The step 4 specifically includes inputting the test set into the verified demand response dynamic evaluation model to obtain the load value after the demand response, and subtracting the baseline load value from the obtained load value to obtain the demand response potential value.

Images