A method for power prediction optimization of a distributed energy storage system

Abstract

The application discloses a kind of distributed energy storage system power prediction optimization methods, first collect the factor characteristics that influence distributed energy storage system use intensity, construct distributed energy storage system intensity prediction sample data set;The sample data set constructed is preprocessed, including attention enhancement and dimensionality reduction processing;Establish machine learning prediction model based on improved random forest, utilize the sample data set after processing to train prediction model;Classify and identify using the prediction model after training is completed to unlabeled sample, according to the result output by prediction model, the use intensity of distributed energy storage system is predicted;Again, according to the use intensity of prediction, distributed energy storage system is maintained and optimized.The above-mentioned method is high in accuracy, strong in robustness, strong in generalization, and can better predict the use intensity of energy storage system, so as to plan power optimization strategy in advance, realize the preposition optimization of energy storage system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric vehicle energy storage systems, in particular to a distributed energy storage system power prediction optimization method. BACKGROUND

[0002] At present, the large-scale entry of electric vehicles into the grid is a great challenge to the planning and operation of the power grid, and is also an opportunity to break through the limitations of electric vehicle development. In the energy storage system of electric vehicles, the prediction of charging power can further study the coordinated charging of electric vehicles, and through the demand side response, the random characteristics of electric vehicle charging are macroscopically scheduled. On the basis of meeting the power demand of electric vehicles, the research on orderly charging strategy is carried out. Electric vehicles, as decentralized energy storage devices, play an important role in the efficient, reliable and safe operation of smart grids, and are an effective way to achieve low-carbon environmental goals and manage smart grid demand. Therefore, using large-scale electric vehicle energy storage characteristics to participate in power distribution network regulation and control, realizing stable operation of the power grid and improving overall economy, has important guiding significance for the optimization of energy storage systems.

[0003] The factors affecting the electric vehicle energy storage system are very complex, and the use behavior of the energy storage system has great randomness in time and space. Although there are various ways to realize the scheduling optimization of the energy storage system in the prior art, there is no solution to predict the use intensity of the energy storage system. SUMMARY

[0004] The purpose of the present application is to provide a distributed energy storage system power prediction optimization method, which has high accuracy, strong robustness and strong generalization ability, and can better predict the use intensity of the energy storage system, so as to plan power optimization strategies in advance and realize the pre-optimization of the energy storage system.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] A distributed energy storage system power prediction optimization method, the method comprising:

[0007] Step 1, collect the factor characteristics affecting the use intensity of the distributed energy storage system, and construct a distributed energy storage system intensity prediction sample data set; wherein the factor characteristics affecting the use intensity of the distributed energy storage system include: electric vehicle scale, charging facility, travel date type, charging price, traffic condition, vehicle performance, and personal subjective travel habit characteristics of the vehicle owner;

[0008] Step 2, pre-process the constructed sample data set, including attention enhancement and dimensionality reduction processing;

[0009] Step 3, establish a machine learning prediction model based on an improved random forest, and train the prediction model using the sample data set processed in step 2.

[0010] Step 4, using the trained prediction model in step 3 to classify and identify unlabeled sample data, and predicting the use intensity of the distributed energy storage system according to the output result of the prediction model;

[0011] Step 5, maintaining and optimizing the distributed energy storage system according to the predicted use intensity.

[0012] As can be seen from the technical solutions provided by the above-mentioned application, the above-mentioned method has high accuracy, strong robustness and strong generalization ability, and can better predict the use intensity of the energy storage system, thereby planning power optimization strategies in advance and realizing pre-optimization of the energy storage system. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor.

[0014] Figure 1 The method flowchart of the power prediction and optimization of the distributed energy storage system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments, which do not constitute a limitation to the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0016] As Figure 1 The method flowchart of the power prediction and optimization of the distributed energy storage system provided by the embodiment of the present application is shown, and the method comprises:

[0017] Step 1, collecting the factor characteristics affecting the use intensity of the distributed energy storage system, and constructing a distributed energy storage system intensity prediction sample data set;

[0018] The factor characteristics affecting the use intensity of the distributed energy storage system include: electric vehicle scale, charging facility, travel date type, charging price, traffic condition, vehicle performance, and personal subjective travel habit characteristics of the vehicle owner;

[0019] In this step, the sample data set is preprocessed, and the specific process is as follows:

[0020] For the character type features of the sample data set, one-hot encoding is used for numerical operation, and the factor features affecting the use intensity of the distributed energy storage system are encoded into feature vectors of the sample data set; wherein the one-hot encoding is a general bag-of-words model for feature extraction, which can encode text and character type data into feature vectors;

[0021] In order to avoid the influence of the dimension between the feature vectors, MinMaxScaler() function is used to normalize the data, so that the preprocessed data is limited in [0, 1];

[0022] Then the processed data is divided into training set and test set, and the training set is labeled; wherein the training set is used for training the subsequent model, and the test set is used for verifying the accuracy of the model.

[0023] Step 2, pre-processing the constructed sample data set, including attention enhancement and dimension reduction processing;

[0024] In this step, the process of attention enhancement processing of the constructed sample data set is as follows:

[0025] Suppose the constructed sample data set is X = [X1, X2, X3,... X n ], the attention mechanism is applied to the calculation, a query vector q is cited, then the association between the model input X and the query vector q is calculated through the scoring function S(x, q), and a score result is output;

[0026] Then the output score result is input into the softmax function for normalization mapping processing, and the corresponding attention distribution a = [a1, a2, a3,... a n ] of the query vector q on the model input X is obtained after processing, and the calculation formula of the attention distribution is as follows:

[0027]

[0028] Wherein, t ∈ [1, 2, 3,... n]; the scoring function S(x, q) is calculated by the following method:

[0029] s(x, q) = v T tanh(Wx+Uq)

[0030] Wherein, W, U are parameters that can be learned and adjusted in the training process, which are set by human and take values in the interval [0, 1];

[0031] Finally, through the attention distribution a = [a1, a2, a3,... a nAttention selection is applied to the model's input, assigning higher weights to important information and lower weights to less important information. This process is iterated continuously to obtain attention-enhanced sample data D, where D = {D1, D2, ..., D...}. n}, in sample data D i The formula for obtaining it is as follows:

[0032] D i =a t ·x i

[0033] Among them, a t For attention distribution; x i Let i be the i-th input sample, i∈[1,2,3,...,n].

[0034] Furthermore, the specific process of dimensionality reduction for the attention-enhanced sample dataset is as follows:

[0035] The number of samples in the attention-enhanced data D is n, and assuming the sample space is p-dimensional, then the data matrix is ​​as follows:

[0036]

[0037] The data is standardized using the following formula:

[0038]

[0039] In the formula: x ij The data for the j-th dimension variable; s is the arithmetic square root of the j-th dimension; j Let be the standard deviation of the j-th dimension of data;

[0040] Then, calculate the correlation coefficient matrix R of the samples:

[0041] R = (r ij ) n×p

[0042] r ij The correlation coefficient between the i-th and j-th dimensions is calculated using the following formula:

[0043]

[0044] In this formula, the asterisk in the upper right corner represents a set, and the asterisk in the upper left corner represents the distance between the current point and the point reconstructed based on the sample.

[0045] Calculate and solve for the eigenvalues ​​of the correlation coefficient matrix R:

[0046] λ1, λ2, ...λ p(lambda1> lambda2> lambda p );

[0047] and the corresponding feature vectors a1, a2, … a p ;

[0048] Then the contribution rate of each eigenvalue is calculated, the first m eigenvalues with cumulative contribution rate greater than 0.85 are taken as the basis for selecting important features, the value of m is set autonomously according to actual conditions, and the calculation formula of the contribution rate is:

[0049]

[0050] Wherein, lambda represents the eigenvalue of the correlation coefficient matrix R; p is the feature serial number;

[0051] The calculation formula of the cumulative contribution rate is:

[0052]

[0053] The first m eigenvalues selected are used to construct a mapping matrix W, and the p-dimensional sample data is converted into an m-dimensional matrix Y in the mapping space, Y = {y1, y2, …, y m}, wherein:

[0054]

[0055] In the formula: a mp is the p-dimensional feature vector corresponding to the mth eigenvalue in the correlation coefficient matrix R; y m is the mth vector of the screened feature.

[0056] Step 3, an improved random forest-based machine learning prediction model is established, and the sample data set processed in step 2 is used to train the prediction model;

[0057] In this step, the traditional random forest does not distinguish all the decision trees, so that the decision trees with different classification performances have the same voting weight. In order to improve the recognition ability of the random forest to the minority class samples, an improved random forest algorithm is proposed in the embodiment of the application. Specifically:

[0058] Based on the traditional random forest decision tree, in the decision tree training stage, the classification performance of each decision tree is evaluated, a higher weight is given to the decision tree which can accurately classify the minority class samples, and the final prediction result is obtained through weighted voting. The prediction result H(T) of the improved random forest is defined as:

[0059]

[0060] Wherein: T is the number of decision trees; I(·) is an indicator function; ft (N) is the prediction result of the tth decision tree; y represents the category; ω t is the voting weight of the tth decision tree.

[0061] When the prediction result of the decision tree is true, the value of the indicator function I(·) is 1, and vice versa.

[0062] The harmonic mean F1 of the precision P re and the recall R ec of each decision tree in classifying the unstable samples is taken as the weight of the decision tree, and the voting weight ω t of each decision tree is defined as:

[0063]

[0064] wherein, TP represents that the stable samples are judged as stable samples, FN represents that the stable samples are judged as unstable samples, FP represents that the unstable samples are judged as stable samples, and TN represents that the unstable samples are judged as unstable samples.

[0065] The greater the harmonic mean F1, the better the classification performance of the decision tree on the minority class samples. The sample data after dimensionality reduction is input into the prediction model based on the improved random forest, and H(T) obtained is the prediction result of the prediction model.

[0066] Step 4, using the prediction model trained in step 3 to classify and identify the unlabeled sample data, and predicting the use intensity of the distributed energy storage system according to the output result of the prediction model;

[0067] In this step, the process of predicting the use intensity of the distributed energy storage system according to the output result of the prediction model is specifically:

[0068] By dividing the data interval, if the output result of the prediction model is located in the abnormal data interval, for example, 1-100, it is predicted that the distributed energy storage system needs to be maintained and optimized in time;

[0069] If the output result of the prediction model is located in the buffer data interval, for example, 101-200, it is predicted that the distributed energy storage system needs to be frequently concerned;

[0070] If the output result of the prediction model is located in the normal data interval, for example, 201-300, it is predicted that the distributed energy storage system does not need to be maintained and optimized.

[0071] Step 5, maintaining and optimizing the distributed energy storage system according to the predicted use intensity.

[0072] In this step, the operation of maintenance and optimization can be carried out by artificial or equipment repair.

[0073] It is worth noting that the content not described in detail in the embodiments of the present application belongs to the prior art known to those skilled in the art.

[0074] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements within the technical range disclosed in the present application can be easily thought by those skilled in the art, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims. The information disclosed in the background section of this document is only intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes the prior art known to those skilled in the art.

Claims

1. A method for distributed energy storage system power prediction optimization, the method comprising: The method comprises: Step 1, collecting factor characteristics affecting the use intensity of the distributed energy storage system, and constructing a distributed energy storage system intensity prediction sample data set; wherein the factor characteristics affecting the use intensity of the distributed energy storage system comprise: electric vehicle scale, charging facility, travel date type, charging price, traffic condition, vehicle performance, and vehicle owner personal subjective travel habit characteristics; Step 2, preprocessing the constructed sample data set, including attention enhancement and dimension reduction processing; Step 3, establishing an improved random forest-based machine learning prediction model, and training the prediction model using the sample data set processed in step 2; The process of step 3 is specifically: Based on the traditional random forest decision tree, in the decision tree training stage, by evaluating the classification performance of each decision tree, a higher weight is given to the decision tree which can accurately classify the minority class samples, and the final prediction result is obtained by weighted voting, which improves the prediction result of random forest is defined as: ； wherein: is the number of decision trees; is the indicator function; is the prediction result of the tth decision tree; denotes the class; is the voting weight of the tth decision tree; When the prediction result of the decision tree is true, the value of the indicator function is 1, and otherwise is 0; the precision of each decision tree in classifying the unstable samples the harmonic mean of precision and recall the voting weight of each decision tree as the weight of the decision tree is defined as: ； ； ； Wherein, TP represents that the stable sample is judged as a stable sample; FN represents that the stable sample is judged as an unstable sample, FP represents that the unstable sample is judged as a stable sample, and TN represents that the unstable sample is judged as an unstable sample; The harmonic mean value The greater, the better the classification performance of the decision tree for the minority class samples, and then input the sample data after dimensionality reduction into the improved random forest-based prediction model to obtain That is, the prediction result of the prediction model; Step 4, classifying and identifying the unlabeled sample data using the prediction model trained in step 3, and predicting the use intensity of the distributed energy storage system according to the output result of the prediction model; Step 5, maintaining and optimizing the distributed energy storage system according to the predicted use intensity.

2. The method of distributed energy storage system power prediction optimization of claim 1, wherein, In step 2, the sample data set is preprocessed, and the specific process is: For the character type characteristics of the sample data set, one-hot encoding is used for numerical operation, and the factor characteristics affecting the use intensity of the distributed energy storage system are encoded into feature vectors of the sample data set; In order to avoid the influence of the dimensions of the feature vectors, the MinMaxScaler() function is used to normalize the data, so that the preprocessed data is limited within [0, 1]; Then the processed data is divided into a training set and a test set, and the training set is labeled; wherein the training set is used for subsequent model training, and the test set is used for verifying the accuracy of the model.

3. The method of claim 1, wherein, In step 2, the process of attention enhancement processing of the sample data set is specifically: Assume a constructed sample dataset is The attention mechanism is applied to the calculation, citing a query vector Then through the scoring function The association between the input of the calculation model And the query vector And output a score result; The output score result is then input to a softmax function for normalized mapping processing, and the query vector is obtained after the processing At the model input The corresponding attention distribution The calculation formula of the attention distribution is as follows: ； wherein ; scoring function is calculated by the method: ； wherein, is a parameter to be learned and adjusted during the training process, and is set artificially to a value in the interval [0, 1]; Finally through the attention distribution Attention selection is performed on the input of the model, higher weights are assigned to important information, lower weights are assigned to unimportant information, and iteration is continuously performed in this process to obtain sample data D based on attention enhancement, , the acquisition formula of the sample data D in is as follows: ； where a t is an attention distribution; x i is a sample of the i-th input, .

4. The method of claim 3, wherein, In step 2, the process of dimension reduction processing of the sample data set after attention enhancement processing is specifically: The sample quantity of the sample data D after the attention enhancement processing is , assuming that the sample space is dimensional, the data matrix at this time is: ； The standardization calculation formula is as follows: ； wherein: is the first dimensional data; is the first dimensional data; is the first dimensional data; Then, the correlation coefficient matrix of the samples is calculated : ； For the first The peacekeeping The correlation coefficient of the peacekeeping ； In the formula, the upper right star represents a set, and the upper left star represents the distance between the current point and the sample reconstruction point; Computing the eigenvalues of the correlation matrix R ； and a corresponding feature vector ; Then the contribution rate of each eigenvalue is calculated, and the first several eigenvalues with cumulative contribution rate greater than 85% are selected as the important features. The value of k is set according to the actual situation, and the calculation formula of the contribution rate is as follows: The value of k is set according to the actual situation, and the calculation formula of the contribution rate is as follows:​ ； where λ represents the eigenvalue of the correlation matrix R; p is the eigen-number; The calculation formula of the cumulative contribution rate is: ； Using the selected front Construct a mapping matrix W using eigenvalues, and then use the mapping space to... Dimensionality reduction of sample data to 3D matrix middle, ,in: ； In the formula: is a correlation coefficient matrix the first eigenvalue corresponds to dimensional eigenvector is the first vector of the screened features.

5. The method of distributed energy storage system power prediction optimization of claim 1, wherein, In step 4, the process of predicting the use intensity of the distributed energy storage system according to the output result of the prediction model is specifically: By dividing the data interval, if the output result of the prediction model is located in the abnormal data interval, it is predicted that the distributed energy storage system needs to be maintained and optimized in time; If the output result of the prediction model is located in the buffer data interval, it is predicted that the distributed energy storage system needs to be frequently concerned; If the output result of the prediction model is located in the normal data interval, it is predicted that the distributed energy storage system does not need to be maintained and optimized.

Images