Electric vehicle charging station charging load prediction method, system, terminal and medium

Abstract

The present invention provides a method and system for predicting the charging load of an electric vehicle charging station. First, an original dataset of charging load is constructed, and a group of basic regressors is generated using the LGBM framework. The TPE algorithm is used to search for hyperparameters in the MongDB space for optimization. The Adaboost structure is used to serially optimize the group of basic regressors to form a final charging load prediction model. The final charging load prediction model can predict the charging load within a very short-term scale, and the obtained charging load prediction curve is more in line with the actual situation. At the same time, a corresponding terminal and medium are provided. Based on ensemble learning, the present invention learns data related to the charging load of the charging station to generate a model, which can output predicted charging values within a very short-term time scale, so as to obtain the charging load curve of the future charging station, providing a decision-making basis for the site selection, expansion, operation, and regulation of the charging station.

Description

Technical Field

[0001] The present invention relates to the technical field of charging load prediction for electric vehicle charging stations, and in particular, to a method, system, terminal and medium for predicting the charging load of an electric vehicle charging station based on integrated learning. Background Art

[0002] The limited supply of fossil resources, air pollution, climate change and global warming have prompted the future transportation system to adopt more efficient and sustainable methods. The emergence of electric vehicles (EVs) is considered an effective alternative to maintain urban transportation by reducing dependence on oil and the resulting air pollution. Electric vehicles have significantly increased their market share in the overall market due to their advantages of using batteries, supercapacitors and fuel cells as energy sources, not relying on fossil fuels, and not emitting polluting gases. According to a report by the International Energy Agency (IEA), under its sustainable development scenario, the global electric vehicle inventory is expected to reach 250 million by 2030.

[0003] The charging power of electric vehicles is uncertain and time-varying; the charging time is highly random and time-period aggregated. The charging behavior of electric vehicle users poses challenges to the safe operation and real-time scheduling of the power grid. Accurate prediction of the charging load of electric vehicle charging stations is an important measure to improve the safe and economic operation of charging stations, and is also an important basis for supporting the decision-making of new construction and capacity expansion planning of charging infrastructure. Experts and scholars at home and abroad have conducted a lot of research on the prediction of electric vehicle charging load, and the main methods are as follows. Based on CGAN adversarial learning, the connection between complex non-linear data is reduced, the deviation of eigenvalue is reduced, and the prediction accuracy is improved, but the training data only contains single-type data and lacks generalization ability. Based on a denoising autoencoder, singular spectrum analysis and long short-term memory neural network to construct a comprehensive power prediction model, the noise interference is reduced and the prediction accuracy is improved, but the model is relatively complex and has high requirements for platform computing resources. Based on the BP-ANN neural network to extract data feature vectors, the CNN neural network to extract image features, and short-term prediction is carried out through multiple layers of BP-ANN neural networks, but the model training time is too long and it is difficult to deploy. Based on the combination of improved random forest and density clustering, the load prediction value is obtained by superimposing the predicted values of each component, but the model construction is complex and the generalization ability is low. Based on the road model, the probability density function of EV charging load is predicted on the travel chain through the Dijkstra path optimization algorithm, but the time scale is long and it cannot face ultra-short-term load prediction. Summary of the Invention

[0004] In view of the above deficiencies in the prior art, the present invention provides a method, system, terminal and medium for predicting the charging load of an electric vehicle charging station based on integrated learning, which is particularly suitable for providing ultra-short-term (i.e., the control scale is within one hour) charging load prediction for an operating charging station.

[0005] According to one aspect of the present invention, there is provided a method for predicting the charging load of an electric vehicle charging station, including:

[0006] Construct an original data set of the charging load of the charging station, and divide the original data set into different weights as the input data set;

[0007] Generate a basic regressor using the input data set, construct a hyperparameter space for the basic regressor, and perform parallel optimization on the selected hyperparameters to obtain a group of basic regressors;

[0008] Select the basic regressors with the largest performance of the loss function in the group of basic regressors for serial integration to obtain an integrated regressor model; construct a hyperparameter space for the regressor model and perform parallel optimization on the selected hyperparameters to obtain a charging load prediction model;

[0009] Use the charging load prediction model to generate future charging load prediction values under a given time scale for predicting the charging load of the charging station.

[0010] Optionally, the construction of the original data set of the charging load of the charging station includes:

[0011] Perform data imputation, variable transformation, and perturbation addition operations on the actual load data, load impact condition data, and random noise data to construct the original data set.

[0012] Optionally, the original data set includes:

[0013] Historical load data Te as the actual load data;

[0014] Charging time T, connection time Tc, charging power P, maximum power MP, start charging time Ts, cut-off charging time Tp, vehicle identification code Tr, charging pile code CP, meteorological condition value, and holiday value as the load impact condition data.

[0015] Optionally, the generation of the basic regressor using the input data set, the construction of the hyperparameter space for the basic regressor, and the parallel optimization of the selected hyperparameters include:

[0016] Select the LGBM framework to generate a basic regressor for the input data set;

[0017] Construct a hyperparameter space with a MongoDB structure for each basic regressor;

[0018] Use the TPE algorithm to find suitable hyperparameters as the selected hyperparameters for parallel optimization.

[0019] Optionally, the selected LGBM framework generates base regressors for the input data set, including:

[0020] Select the LGBM framework, use the input data set as input, and learn from the input data set according to the LGBM framework to obtain multiple base regressors for generating a base regressor group;

[0021] Determine the splitting point of the input data set according to the variance gain

[0022]

[0023] In the formula: n is the number of features in the training data set, j is the feature selected in the algorithm, x i is the input space of the data, d is the splitting point of the data, A l is a part of the training data with a larger gradient performance, B l is A l The data set constructed by sampling the remaining part, g i is the gradient space, a is A l and the remaining part of A r The smaller part extracted from the remaining part, b is A l and the remaining part of A r The larger part extracted from the remaining part, is the feature in the gradient extracted from the l part at the splitting point d, A r is A l The part with a smaller remaining gradient performance, B r is A r The data set constructed by sampling the remaining part, is the feature in the gradient extracted from the r part at the splitting point d;

[0024] The approximate error ε(d) of the base regressor is obtained as:

[0025]

[0026] In the formula: C a,b is the largest data point in the normalized gi gradient space, δ is the probability of data point selection, and D is the largest gradient value at the splitting point d;

[0027] Constructing a hyperparameter space with a MongoDB structure for each of the base regressors, and using the TPE algorithm to find suitable hyperparameters as the selected hyperparameters for parallel optimization, including:

[0028] For the generated initial group of basic regressors, select their learning rate, maximum growth depth, number of iterations, number of leaf nodes, subsampling instances, L1 regularization, and L2 regularization parameters as the construction categories of hyperparameters; based on the MongoDB architecture, construct a hyperparameter search space with a MongoDB structure;

[0029] Based on the hyperparameter search space, use the TPE algorithm to obtain hyperparameters that meet the conditions for parallel optimization; construct models using both the lower and higher values in the same category of hyperparameters, and use the loss function SMAPE to evaluate the performance of the models respectively, where the loss function SMAPE is expressed as follows:

[0030]

[0031] In the formula: n is the maximum number of basic regressors, A t is the existing performance of the generated model, F t is the predicted performance of the model;

[0032] Calculate the SMAPE value of the model constructed by the hyperparameters, and select the hyperparameter group whose SMAPE value is greater than the 1 / 2 quantile of the SMAPE of all basic regressors for the next round of iteration. After continuously selecting hyperparameters for the process of building-evaluating-optimizing, output the optimized group of basic regressors with the optimal selected hyperparameter combination.

[0033] Optionally, select the group of basic regressors with the largest SMAPE performance of the loss function in the group of basic regressors for serial integration to obtain an integrated regressor model; construct a hyperparameter space for the regressor model and perform parallel optimization on the selected hyperparameters, including:

[0034] Use the Adaboost algorithm to select the group of basic regressors with the largest loss function performance in the group of basic regressors for serial integration to obtain an integrated regressor model;

[0035] For the integrated regressor model, select the number of basic regressors, the weight reduction coefficient of the basic regressors, and the type of error function as the construction types of hyperparameters, and based on the MongoDB architecture, construct a space with a MongoDB structure for storing hyperparameter values; use the TPE algorithm to obtain hyperparameters that meet the conditions as the selected hyperparameters and perform parallel optimization.

[0036] Optionally, the use of the Adaboost algorithm to select the group of basic regressors with the largest SMAPE performance of the loss function in the group of basic regressors for serial integration includes:

[0037] Calculate the weighted regression error rate ∈ of each basic regressor in the group of basic regressors m :

[0038]

[0039] Wherein, G m (x i ) is an arbitrarily selected base regressor, is the initial weight of a certain base regressor; y i is the actual predicted result, II is the indicator function, when the expression in the parentheses is true, the value of the whole expression is 1, and when the expression in the parentheses is false, the value of the whole expression is 0, and N is the maximum number of base regressors;

[0040] Calculate the weight coefficient α m (x i ) of the arbitrarily selected base regressor G in the final regressor m :

[0041]

[0042] According to the weighted regression error rate ∈ m Change the weight coefficient α m , update the weight distribution of the input data set, obtain the new weight distribution of the data set, and use it for the next iteration;

[0043] After the above steps, the serial integration of the base regressor group is completed.

[0044] According to the second aspect of the present invention, a charging load prediction system for an electric vehicle charging station is provided, including:

[0045] An input data set construction module, which constructs the original data set of the charging load of the charging station and divides the original data set into different weights as the input data set;

[0046] A base regressor group construction module, which uses the input data set to generate base regressors, constructs a hyperparameter space for the base regressors, and performs parallel optimization on the selected hyperparameters to obtain a base regressor group;

[0047] A charging load prediction model construction module, selects the base regressor group with the largest loss function performance in the base regressor group for serial integration to obtain an integrated regressor model; constructs a hyperparameter space for the regressor model, and performs parallel optimization on the selected hyperparameters to obtain a charging load prediction model;

[0048] A very short-term prediction module, which uses the charging load prediction model to generate the future charging load prediction value under a given time scale for the very short-term prediction of the charging load of the charging station.

[0049] According to a third aspect of the present invention, there is provided a terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it can be used to execute the method described in any one of the above, or run the above system.

[0050] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it can be used to execute the method described in any one of the above when the program is executed, or run the above system.

[0051] Due to the adoption of the above technical solutions, compared with the prior art, the present invention has at least one of the following beneficial effects:

[0052] The electric vehicle charging station charging load prediction method, system, terminal and medium provided by the present invention, based on the integration idea, first constructs an original data set of charging load, generates a group of basic regressors using the LGBM framework, uses the TPE algorithm to search for hyperparameters in the MongDB space for optimization, and uses the Adaboost structure to serially optimize the group of basic regressors to form a final charging load prediction model. The final charging load prediction model can predict the charging load within a very short-term scale, and the obtained charging load prediction curve is more in line with the actual situation.

[0053] The electric vehicle charging station charging load prediction method, system, terminal and medium provided by the present invention overcomes the problems of large volume, high timeliness, and high computing resources of the data for ultra-short-term charging load prediction. Based on ensemble learning, it learns the data related to the charging load of the charging station to generate a model, and this model can output the predicted charging value within an ultra-short-term time scale, thereby obtaining the charging load curve of the future charging station, providing a decision-making basis for the site selection, expansion, operation and regulation of the charging station.

[0054] The electric vehicle charging station charging load prediction method, system, terminal and medium provided by the present invention learns the original data set through the LGBM algorithm in the first layer to construct a group of basic regressors; in the second layer, it serially constructs the group of basic regressors in the first step in the Adaboost mode, and finally integrates them into a charging load prediction model.

[0055] The electric vehicle charging station charging load prediction method, system, terminal and medium provided by the present invention can have a very low time scale for predicting the charging load value of electric vehicles in the prediction area, thereby realizing the charging load prediction within a very short-term scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] By reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, other features, objects and advantages of the present invention will become more apparent:

[0057] Figure 1 This is the flowchart of the charging load prediction method for an electric vehicle charging station in an embodiment of the present invention.

[0058] Figure 2 This is the hyperparameter space structure for storing hyperparameters established in a preferred embodiment of the present invention.

[0059] Figure 3 This is the schematic diagram of the working process of the charging load prediction method for an electric vehicle charging station in a preferred embodiment of the present invention.

[0060] Figure 4 This is the schematic diagram of the component modules of the charging load prediction system for an electric vehicle charging station in an embodiment of the present invention. Detailed implementation manners

[0061] The following is a detailed description of the embodiments of the present invention: These embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

[0062] Figure 1 This is the flowchart of the charging load prediction method for an electric vehicle charging station provided in an embodiment of the present invention.

[0063] As Figure 1 shown, the charging load prediction method for an electric vehicle charging station provided in this embodiment may include the following steps:

[0064] S100. Construct an original data set of the charging load of the charging station, and divide the original data set into different weights as the input data set;

[0065] S200. Use the input data set to generate a basic regressor, construct a hyperparameter space for the basic regressor, and perform parallel optimization on the selected hyperparameters to obtain a group of basic regressors;

[0066] S300. Select the basic regressors with the largest performance of the loss function in the group of basic regressors for serial integration to obtain an integrated regressor model; construct a hyperparameter space for the regressor model, and perform parallel optimization on the selected hyperparameters to obtain a charging load prediction model;

[0067] S400. Use the charging load prediction model to generate future charging load prediction values under a given time scale for predicting the charging load of the charging station.

[0068] In S100 of this embodiment, as a preferred embodiment, constructing the original data set of the charging load of the charging station may include the following steps:

[0069] Perform data imputation, variable transformation, and perturbation addition operations on the actual load data, load impact condition data, and random noise data to construct the original dataset.

[0070] In S100 of this embodiment, as a preferred embodiment, the original dataset includes:

[0071] Historical load data Te as the actual load data;

[0072] Charging time T, connection time Tc, charging power P, maximum power MP, start charging time Ts, cut-off charging time Tp, vehicle identification code Tr, charging pile code CP, meteorological condition values, and holiday values as the load impact condition data.

[0073] In S200 of this embodiment, as a preferred embodiment, using the input dataset to generate a base regressor, constructing a hyperparameter space for the base regressor, and performing parallel optimization on the selected hyperparameters may include the following steps:

[0074] S201, select the LGBM framework to generate a base regressor for the input dataset;

[0075] S202, construct a hyperparameter space with a MongoDB structure for each base regressor;

[0076] S203, use the TPE algorithm to find suitable hyperparameters as the selected hyperparameters for parallel optimization.

[0077] In S201 of this embodiment, as a preferred embodiment, selecting the LGBM framework to generate a base regressor for the input dataset may include the following steps:

[0078] S2011, select the LGBM framework, use the input dataset as the input, and learn from the input dataset according to the LGBM framework to obtain multiple base regressors for generating a base regressor group;

[0079] S2012, determine the split point of the input dataset according to the variance gain

[0080]

[0081] Where: n is the number of features in the training dataset, j is the feature selected in the algorithm, x i is the input space of the data, d is the split point of the data, A l is a part of the training data with a greater gradient performance, B l is the dataset constructed by sampling the remaining part of A l g iis the gradient space, and a is A l and A r a smaller part extracted from the remaining part, and b is A l and A r a larger part extracted from the remaining part is the feature in the gradient extracted from the l part at the splitting point d, A r is A l a part where the remaining gradient shows a smaller value, B r is A r a dataset constructed by sampling the remaining part is the feature in the gradient extracted from the r part at the splitting point d;

[0082] In S2013, the approximate error of the base regressor is obtained as follows:

[0083]

[0084] In the formula: C a,b is the normalized g i the largest data point in the gradient space, δ is the probability of data point selection, and D is the largest gradient value at the splitting point d;

[0085] In S202 and S203 of this embodiment, as a preferred embodiment, for each base regressor, a hyperparameter space with a MongoDB structure is constructed, and the TPE algorithm is used to find suitable hyperparameters as the selected hyperparameters for parallel optimization, which may include the following steps:

[0086] For the generated initial base regressor group, select its learning rate, maximum growth depth, number of iterations, number of leaf nodes, subsampling instances, L1 regularization, and L2 regularization parameters as the construction categories of hyperparameters; based on the MongoDB architecture, construct a hyperparameter search space with a MongoDB structure;

[0087] Based on the hyperparameter search space, use the TPE algorithm to obtain hyperparameters that meet the conditions for parallel optimization; use the lower and higher values of the same type of hyperparameters to construct models at the same time, and use the loss function SMAPE to evaluate the performance of the models respectively, where the loss function SMAPE is expressed as follows:

[0088]

[0089] In the formula: n is the maximum number of base regressors, A t is the existing performance of the generated model, F t is the expected performance of the model;

[0090] Calculate the SMAPE value of the model constructed by the hyperparameters, and select the hyperparameter group whose SMAPE value is greater than the 1 / 2 quantile of the SMAPE of all base regressors for the next iteration. After continuously selecting hyperparameters for the process of establishment-evaluation-optimization, output the optimized base regressor group with the optimal selected hyperparameter combination.

[0091] Further, in a specific application example, based on the MongoDB architecture, construct a hyperparameter search space with a MongoDB structure, including the following steps:

[0092] For the learning rate, the value range is 0.05 - 0.1, the maximum growth depth is 3 - 5, the number of iterations is 100 - 1000, the number of leaf nodes is 7 - 9, the subsampling instance is 0.7 - 1.0, L1 regularization is 0 - 1000, and L2 regularization is 0 - 1000;

[0093] The values of the hyperparameters of each category are arranged in ascending order in the data storage part of MongDB. The values below 100 are stored in the low-value part, and the values above 100 are stored in the high-value part;

[0094] In the architecture management of MongoDB, first back up all the value data and perform automatic splitting;

[0095] Next, construct an index for the class map of all data hyperparameters and number each value; Draw on the four operating characteristics of the MongoDB structure: 1) Dynamic query: Consider the available index and directly read the content from the cache and perform addition, deletion, query, and modification of data; 2) Index: In the MongoDB class structure, an id number is created for each piece of data as the most basic index; 3) Geographical location index: The index number can be based on location-related attributes such as distance; 4) Pre-query: Before actually searching for a piece of data, test the time-consuming of the search operation as a detection of the database search efficiency;

[0096] Thus, construct a hyperparameter search space with a MongoDB structure.

[0097] In S300 of this embodiment, as a preferred embodiment, select the base regressor group with the largest SMAPE value of the loss function in the base regressor group for serial integration to obtain an integrated regressor model; construct a hyperparameter space for the regressor model and perform parallel optimization on the selected hyperparameters, which may include the following steps:

[0098] S301, Use the Adaboost algorithm to select the base regressor group with the largest loss function performance in the base regressor group for serial integration to obtain an integrated regressor model;

[0099] S302. For the integrated regressor model, select the number of base regressors, the weight reduction coefficient of the base regressors, and the type of error function as the construction types of hyperparameters. Based on the MongoDB architecture, construct a MongoDB structure for storing the hyperparameter value space; use the TPE algorithm to obtain the hyperparameters that meet the conditions as the selected hyperparameters and perform parallel optimization.

[0100] Further, in a specific application example, in the method of constructing a MongoDB structure for storing the hyperparameter value space based on the MongoDB architecture, the number of base regressors is set to 50 - 100, the weight reduction coefficient of the base regressors is 0 - 1, and the types of error functions are linear, sum of squares, and exponential. The architecture and structure adopted in this method are the same as those in the hyperparameter search space for constructing the MongoDB structure, which will not be elaborated here.

[0101] In S301 of this embodiment, as a preferred embodiment, using the Adaboost structure to select the base regressor group with the largest loss function performance in the base regressor group for serial integration may include the following steps:

[0102] S3011. Calculate the weighted regression error rate ∈ of each base regressor in the base regressor group m :

[0103]

[0104] In the formula, G m (x i ) is an arbitrarily selected base regressor, is the initial weight of a certain base regressor; y i is the actual predicted result, II is the indicator function, when the expression in the parentheses is true, the value of the whole expression is 1, and when the expression in the parentheses is false, the value of the whole expression is 0, and N is the maximum number of base regressors;

[0105] S3012. Calculate the weight coefficient α m (x i ) of an arbitrarily selected base regressor G in the final regressor m :

[0106]

[0107] S3013. According to the weighted regression error rate ∈ m change the weight coefficient α m , update the weight distribution of the input data set to obtain the new weight distribution of the data set for the next iteration.

[0108] After the above steps, the serial integration of the basic regressor group is completed.

[0109] The following further details the technical solutions provided in the above embodiments of the present invention with reference to the accompanying drawings.

[0110] Figure 2 It is a schematic diagram of the hyperparameter space structure constructed based on the MongoDB format. MongoDB is a distributed storage format. For example, Figure 2 As shown, the above embodiments of the present invention use the MongoDB format to construct the hyperparameter space. Instead of verifying only one combination of hyperparameters each time, all combinations of hyperparameters can be stored in the MongoDB format, making these combinations not easily lost and easier to find.

[0111] Figure 3 It is a working schematic diagram of the charging load prediction method for an electric vehicle charging station provided by a preferred embodiment of the present invention.

[0112] For example, Figure 3 As shown, the charging load prediction method for the electric vehicle charging station provided by this preferred embodiment selects the charging load of the charging station as the training subject and selects a new energy charging station in the Netherlands as the load prediction area. This area includes residential areas, commercial office buildings, and shopping malls, with a large number of moving vehicles and four types of electric vehicle charging clusters. The charging load prediction method for the electric vehicle charging station is based on ensemble learning and includes the following steps:

[0113] I. Dataset construction

[0114] First, construct the original dataset of the charging load of the charging station. The actual load data, load impact condition data, and random noise data are subjected to data imputation, variable transformation, and perturbation addition operations to construct the original dataset, and different weights are assigned to the original dataset as the input for the next layer.

[0115] II. Generate basic regressors

[0116] Select the LGBM framework to generate basic regressors for the original dataset with the original weights divided in step I. Construct a hyperparameter space with a MongoDB structure for each basic regressor, and use the TPE algorithm to find appropriate hyperparameters for parallel optimization.

[0117] III. Serial integration of the basic regressor group

[0118] Use the Adaboost structure to select the basic regressor group with the largest loss function performance in the generated basic regressor group for serial integration. Construct a hyperparameter space with a MongoDB structure for the finally optimized integrated model, and use the TPE algorithm to find appropriate hyperparameters for parallel optimization.

[0119] IV. Output the load prediction value of the electric vehicle charging station

[0120] The finally integrated and optimized model can generate the future charging load prediction value under a given time scale for the ultra-short-term prediction of the charging load of the charging station.

[0121] In this preferred embodiment, in step one, the original data set includes historical load data Te, charging time T, connection time Tc, charging power P, maximum power MP, start charging time Ts, cut-off charging time Tp, vehicle identification code Tr, charging pile code CP, meteorological condition values, and holiday values.

[0122] Random noise data is randomly added in the experiment. In this preferred embodiment, in step two, the process of generating the base regressor is as follows:

[0123] 2.1. Using the data set with the original weights divided in step one as the input, learn the data set according to the LGBM framework to obtain several base regressors. Determine the split point of the training data according to the variance gain, and its formula is:

[0124]

[0125] In the formula: n is the number of features in the training data set, j is the feature selected in the algorithm, x i is the input space of the data, d is the split point of the data, A l is a part of the training data with a larger gradient performance, B l is the data set constructed by sampling the remaining part of A l g i is the gradient space, a is the smaller part extracted from A l and the remaining part of A r b is the larger part extracted from A l and the remaining part of A r The remaining part, is the feature in the gradient extracted from the l part at the split point d, A r is A l The remaining part with a smaller gradient performance, B r is A r The data set constructed by sampling the remaining part, is the feature in the gradient extracted from the r part at the split point d;

[0126] The split point is used to finally find the best division point in the global histogram.

[0127] For feature j, the decision tree algorithm selects and calculates the maximum gain V j (d). Then according to feature j *At the point The data is split into left and right child nodes. They are sorted in descending order according to the absolute value of the gradient of the training data; and a×100% of the data and the larger gradients in the head part are retained to obtain a data subset A; then, for the remaining subset A c set, which consists of (1 - a)×100% of the data with smaller gradients, and then a subset B of size b×|A c | is further randomly sampled; finally, according to the estimated variance gain V j (d) Split the data on the subset A∪B. Used to normalize the gradients on B back to the size of A c .

[0128] The approximation error of the base regressor is:

[0129]

[0130] Furthermore:

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] In the formula: C a,b is the largest data point in the normalized g i gradient space, δ is the probability of data point selection, and D is the largest gradient value at the splitting point d;

[0138] 2.2. For the generated base regressor group framework, parameters such as the maximum growth depth, the number of leaf nodes, the subsampling instance, the random partial features, L1 regularization, and L2 regularization are selected to construct a parameter search space in the MongoDB mode, and the TPE algorithm is selected for parallel optimization to output an optimized base regressor group with the optimal selected hyperparameter combination.

[0139] In this preferred embodiment, in step three, the base regressor group is optimized and integrated in a serial manner with the Adaboost structure, and the integration process is as follows:

[0140] 3.1. Calculate the weighted regression error rate ∈ m of each base regressor in the base regression group generated in step two, and its calculation formula is:

[0141]

[0142] Among them, G m (x i ) is an arbitrarily extracted basic regressor, is the initial weight of a certain basic regressor.

[0143] 3.2. Calculate the weight coefficient α m of G(x) in the final regressor, m and its calculation formula is:

[0144]

[0145] Change the weight coefficient according to the regression error rate, update the weight distribution of the training data set, obtain the new weight distribution of the samples, and use it for the next round of iteration.

[0146] The basic regressor group is vulnerable to the bias-variance trade-off. Using the Adaboost algorithm for integration can increase the prediction ability of the model while maintaining the balance of the model variance and bias, and avoid problems such as the reduction of regression accuracy and the distortion of results when a single model is too dependent on the training set during test set and real prediction, and finally obtain a stronger and better model.

[0147] In this preferred embodiment, the way of ensemble learning in step three is selected according to the strength of R 2 performance. Among them, after selecting the basic regressors, there are 4 ways to integrate the basic regressors: Boosting, Bagging, Blending, and Stacking. By comparison, Adaboost in the best-performing Boosting category is selected to integrate all basic regressors.

[0148] In this preferred embodiment, in step four, whether the charging load prediction model is satisfied is judged by integrating the following two parts:

[0149] 4.1. Learn the original data set through the LGBM algorithm, generate basic regressors, and construct them into a basic regressor group.

[0150] 4.2. Serially integrate the basic regressor group in the first step in the Adaboost mode, and generate the final charging load prediction model after adjusting and optimizing the hyperparameters of the finally integrated model.

[0151] In this preferred embodiment, in step four, the time scale for predicting the charging load of electric vehicles in the prediction area is extremely low and belongs to the ultra-short-term scale.

[0152] In this preferred embodiment, in step four, the time scale for outputting the predicted value is controllable, and the predicted load curve for future time scales such as one hour or one day can be obtained according to the model output.

[0153] Figure 4 FIG. is a schematic diagram of the composition modules of an ultra-short-term charging load prediction system for an electric vehicle charging station provided by an embodiment of the present invention.

[0154] As Figure 4 shown, the ultra-short-term charging load prediction system for an electric vehicle charging station provided by this embodiment may include the following modules:

[0155] An input data set construction module, which constructs the original data set of the charging load of the charging station and divides the original data set into different weights as the input data set;

[0156] A basic regressor group construction module, which uses the input data set to generate basic regressors, constructs a hyperparameter space for the basic regressors, and performs parallel optimization on the selected hyperparameters to obtain a group of basic regressors;

[0157] A charging load prediction model construction module, which selects the basic regressors with the largest loss function performance in the group of basic regressors for serial integration to obtain an integrated regressor model; constructs a hyperparameter space for the regressor model, and performs parallel optimization on the selected hyperparameters to obtain a charging load prediction model;

[0158] An ultra-short-term prediction module, which uses the charging load prediction model to generate the future charging load prediction value under a given time scale for ultra-short-term prediction of the charging load of the charging station.

[0159] It should be noted that the steps in the method provided by the present invention can be implemented by corresponding modules, devices, units, etc. in the system. Those skilled in the art can refer to the technical solution of the method to implement the composition of the system. That is, the embodiments in the method can be understood as the preferred examples for constructing the system, which will not be elaborated here.

[0160] An embodiment of the present invention provides a terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it can be used to execute the method of any one of the above embodiments of the present invention, or run the system of any one of the above embodiments of the present invention.

[0161] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it can be used to execute the method of any one of the above embodiments of the present invention, or run the system of any one of the above embodiments of the present invention.

[0162] In the above two embodiments, optionally, a memory for storing programs; the memory may include a volatile memory (English: volatile memory), such as a random access memory (English: random-access memory, abbreviation: RAM), such as a static random access memory (English: static random-access memory, abbreviation: SRAM), a double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviation: DDR SDRAM), etc.; the memory may also include a non-volatile memory (English: non-volatile memory), such as a flash memory (English: flash memory). The memory is used to store computer programs (such as application programs and functional modules for implementing the above methods), computer instructions, etc. The above computer programs, computer instructions, etc. can be stored in partitions in one or more memories. And the above computer programs, computer instructions, data, etc. can be called by the processor.

[0163] The above computer programs, computer instructions, etc. can be stored in partitions in one or more memories. And the above computer programs, computer instructions, data, etc. can be called by the processor.

[0164] A processor for executing the computer programs stored in the memory to implement each step in the methods related to the above embodiments. For specific details, reference can be made to the relevant descriptions in the previous method embodiments.

[0165] The processor and the memory can be of independent structures or integrated structures. When the processor and the memory are of independent structures, the memory and the processor can be coupled and connected through a bus.

[0166] The electric vehicle charging station charging load prediction method, system, terminal and medium provided in the above embodiments of the present invention are based on integrated learning. First, the actual load data, influencing condition data, and random noise data related to the charging station load are merged to construct an original data set. Then, LGBM is used to generate a group of base regressors, and an Adaboost structure is adopted to perform a serial integrated prediction model on the group of base regressors. The TPE algorithm is used to optimize the model in the hyperparameter space of the MongoDB structure. Finally, the integrated model performs a very short-term prediction on the charging load. It can output the predicted charging value within a very short-term time scale, thereby obtaining the charging load curve of the future charging station, providing a decision-making basis for the site selection, expansion, operation, and regulation of the charging station.

[0167] Those skilled in the art know that, in addition to implementing the system and its various devices provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the system and its various devices provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same functions. Therefore, the system and its various devices provided by the present invention can be regarded as a kind of hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component; the devices for implementing various functions can also be regarded as either software modules for implementing the method or structures within the hardware component.

[0168] Matters not described in detail in the above embodiments of the present invention are all well-known techniques in the art.

[0169] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for predicting the charging load of an electric vehicle charging station, characterized in that, Including: Constructing an original dataset of the charging load of a charging station, and dividing the original dataset into different weights as an input dataset; Generating a base regressor using the input dataset, constructing a hyperparameter space for the base regressor, and performing parallel optimization on the selected hyperparameters to obtain a group of base regressors; including: Selecting the LGBM framework, using the input dataset as input, and learning from the input dataset according to the LGBM framework to obtain multiple base regressors for generating a group of base regressors; Determining the splitting point of the input dataset according to the variance gain; Obtaining the approximation error of the base regressor; For the generated initial group of base regressors, selecting its learning rate, maximum growth depth, number of iterations, number of leaf nodes, subsampling instances, L1 regularization, and L2 regularization parameters as the construction categories of hyperparameters; based on the MongoDB architecture, constructing a hyperparameter search space with a MongoDB structure; Based on the hyperparameter search space, using the TPE algorithm to obtain hyperparameters that meet the conditions for parallel optimization; constructing models simultaneously using lower and higher values in the same category of hyperparameters, and using the loss function SMAPE to evaluate the performance of the models respectively; Calculating the SMAPE value of the model constructed by the hyperparameters, selecting the hyperparameter group whose SMAPE value is greater than the 1 / 2 quantile of the SMAPE of all base regressors for the next round of iteration. After continuously selecting hyperparameters for establishment, evaluation, and optimization, outputting an optimized group of base regressors with the optimal selected hyperparameter combination; Selecting the group of base regressors with the largest performance of the loss function in the group of base regressors for serial integration to obtain an integrated regressor model; constructing a hyperparameter space for the regressor model and performing parallel optimization on the selected hyperparameters to obtain a charging load prediction model; Using the charging load prediction model to generate future charging load prediction values at a given time scale for predicting the charging load of the charging station.

2. The method for predicting the charging load of an electric vehicle charging station according to claim 1, wherein The construction of the original dataset of the charging load of the charging station includes: Performing data imputation, variable transformation, and adding perturbations on the actual load data, load impact condition data, and random noise data to construct the original dataset.

3. The method for predicting the charging load of an electric vehicle charging station according to claim 2, wherein The original dataset includes: Historical load data Te as the actual load data; Charging time T, connection time Tc, charging power P, maximum power MP, start charging time Ts, end charging time Tp, vehicle identification code Tr, charging pile code CP, meteorological condition values, and holiday values as the load impact condition data.

4. The method for predicting the charging load of an electric vehicle charging station according to claim 1, wherein Determine the split point of the input data set according to the variance gain as follows: Where: n is the number of features in the training dataset, j is the feature selected in the algorithm, x i is the input space of the data, d is the splitting point of the data, A l is a part of the training data with a larger gradient performance, B l is the remaining part of A l The dataset constructed by sampling the remaining part, g i is the gradient space, a is the smaller part extracted from A l and the remaining part of A r b is the larger part extracted from A l and the remaining part of A r The remaining part; is the feature in the gradient extracted from part 1 at the splitting point d, A r is the remaining part of A l with a smaller gradient performance, B r is the remaining part of A r The dataset constructed by sampling the remaining part, is the feature in the gradient extracted from part r at the splitting point d; The approximation error ε(d) of the base regressor is obtained as: Where: C a,b is the maximum data point in the normalized g i gradient space, δ is the probability of data point selection, and D is the maximum gradient value at the segmentation point d; The loss function SMAPE is expressed as follows: Where: n is the maximum number of base regressors, A t is the existing performance of the generative model, F t is the predicted performance of the model.

5. The method for predicting the charging load of an electric vehicle charging station according to claim 1, wherein The selection of the group of base regressors with the largest performance of the loss function in the group of base regressors for serial integration to obtain an integrated regressor model; constructing a hyperparameter space for the regressor model and performing parallel optimization on the selected hyperparameters includes: The Adaboost algorithm is used to select the base regressor group with the largest loss function performance in the base regressor group for serial integration, and an integrated regressor model is obtained; For the integrated regressor model, the number of base regressors, the weight reduction coefficient of the base regressor, and the type of error function are selected as the construction types of hyperparameters. Based on the MongoDB architecture, a MongoDB structure is constructed for the space storing hyperparameter values; the TPE algorithm is used to obtain hyperparameters that meet the conditions as the selected hyperparameters and perform parallel optimization.

6. The method for predicting the charging load of an electric vehicle charging station according to claim 5, wherein The use of the Adaboost algorithm to select the base regressor group with the largest loss function performance in the base regressor group for serial integration includes: Calculate the weighted regression error rate ò of each base regressor in the base regressor group m : Where, G m (x i ) is an arbitrarily extracted basic regressor, is the initial weight of a certain basic regressor; y i is the actual predicted result, is an indicator function. When the expression in the parentheses is true, the value of the entire expression is 1; when the expression in the parentheses is false, the value of the entire expression is 0. N is the maximum number of basic regressors; Calculate the weight coefficient α of the arbitrarily selected base regressor G in the final regressor for the given base regression group m (x i ) m : According to the weighted regression error rate change the weight coefficient α m , update the weight distribution of the input data set to obtain a new weight distribution of the data set for the next iteration; After the above steps, the serial integration of the base regressor group is completed.

7. An electric vehicle charging station charging load prediction system, characterized in that Including: An input dataset construction module, which constructs an original dataset of the charging load of the charging station and divides the original dataset into different weights as the input dataset; A base regressor group construction module, which uses the input dataset to generate base regressors, constructs a hyperparameter space for the base regressors, and performs parallel optimization on the selected hyperparameters to obtain a base regressor group; including: Select the LGBM framework, use the input dataset as the input, and learn the input dataset according to the LGBM framework to obtain multiple base regressors for generating the base regressor group; Determine the splitting point of the input dataset according to the variance gain; Obtain the approximate error of the base regressor; For the generated initial base regressor group, select its learning rate, maximum growth depth, number of iterations, number of leaf nodes, subsampled instances, L1 regularization, and L2 regularization parameters as the construction categories of hyperparameters; based on the MongoDB architecture, construct a hyperparameter search space in the MongoDB structure; Based on the hyperparameter search space, use the TPE algorithm to obtain hyperparameters that meet the conditions for parallel optimization; use the lower and higher values in the same type of hyperparameters to construct models at the same time, and use the loss function SMAPE to evaluate the performance of the models respectively; Calculate the SMAPE value of the model constructed by the hyperparameters, select the hyperparameter group with the SMAPE value greater than the 1 / 2 quantile of the SMAPE of all base regressors for the next round of iteration. After continuously selecting hyperparameters for establishment, evaluation, and optimization, output the optimized base regressor group with the optimal selected hyperparameter combination; A charging load prediction model construction module, which selects the base regressor group with the largest loss function performance in the base regressor group for serial integration to obtain an integrated regressor model; constructs a hyperparameter space for the regressor model and performs parallel optimization on the selected hyperparameters to obtain a charging load prediction model; A very short-term prediction module, which uses the charging load prediction model to generate future charging load prediction values at a given time scale for the very short-term prediction of the charging load of the charging station.

8. A terminal, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it can be used to execute the method described in any one of claims 1-6, or, run the system described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can be used to perform the method described in any one of claims 1-6, or to run the system described in claim 7.

Images