Platform area charging capacity interval prediction method

By collecting and processing station data, using multi-layer neural networks and fuzzy kernel density estimation methods and other technologies, real-time rolling prediction of the charging capacity interval in the station area is achieved, solving the problems of prediction difficulties and high resource consumption in the existing technology, and improving the problem of improvement. The accuracy and efficiency of predictions are achieved.

CN118966812AInactive Publication Date: 2024-11-15SICHUAN UNIV

Patent Information

Application Number
CN202410549614.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-06
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict the available charging capacity in the station area during the charging period of electric vehicles, resulting in difficulty in power balance in the power system, increased operation and maintenance risks of intelligent charging piles, and the prediction process consumes high computing power and memory.

Method used

By collecting monitoring data, meteorological data and calendar information data from the platform area, overall status data is generated, and real-time rolling prediction is performed using the charging capacity interval prediction model. The model includes a multi-layer neural network structure, combined with fuzzy kernel density estimation method and ten-fold cross-validation reduces computing power demand and memory usage.

Benefits of technology

Real-time rolling prediction of the available charging capacity range in the station area with less memory usage and low computing power demand is achieved, which improves the accuracy and efficiency of prediction and reduces operation and maintenance costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966812A_ABST
    Figure CN118966812A_ABST
Patent Text Reader

Abstract

The embodiment of the invention discloses a transformer area charging capacity interval prediction method. The method comprises the steps of collecting monitoring data corresponding to a transformer area, meteorological data corresponding to the transformer area and calendar information data; generating transformer area data based on the monitoring data, the meteorological data and the calendar information data; and through a charging capacity interval prediction model, according to the transformer area data, generating at least one target prediction interval, each target prediction interval being used for representing a charging capacity interval prediction result corresponding to each different prediction confidence. Compared with the prior art, the embodiment of the invention can realize real-time rolling prediction of the available charging capacity interval of the transformer area under the conditions of less memory occupation and lower computing power demand.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method for predicting a charging capacity interval of a station area. Background Art

[0002] The rapid expansion of electric vehicles (EVs) is accompanied by the construction and operation of a large number of public charging infrastructure. The grid connection of charging piles will bring a series of challenges to low-voltage distribution stations, such as: increased system load peaks; overload of cables, wires and transformers; increased network losses; and prominent low-voltage problems at the end. Reasonable operation and maintenance of smart charging piles is the key to ensuring the safe and reliable operation of low-voltage distribution stations. However, the charging period of electric vehicles overlaps with the peak of residential electricity load, and the carrying capacity of low-voltage distribution stations is limited. How to reasonably control the charging power and the charging time process is a major problem faced by operators. Accurate and reliable prediction of the available charging capacity of electric vehicles in the distribution area is of great significance for maintaining the power balance of the power system, improving the large-scale access capacity of electric vehicles, optimizing the operation of the distribution network, safe and stable control of the power grid, determining the carrying capacity of charging piles, building the transportation energy Internet, planning the layout of charging stations in the distribution area, and guiding the orderly charging of electric vehicles.

[0003] However, meteorological conditions, holiday factors, and end-user production and consumption behaviors are highly uncertain, making it difficult to accurately predict the available charging capacity of the substation. Traditional point predictions are difficult to characterize the uncertainty of the substation status, which increases the operating risk of smart charging piles. At the same time, the prediction process consumes high computing power and memory usage, making the cost of predicting the available charging capacity of the substation remain high. Summary of the invention

[0004] The embodiment of the present application provides a method for predicting the charging capacity interval of a substation. Compared with the prior art, the embodiment of the present application can realize real-time rolling prediction of the available charging capacity interval of the substation with less memory occupancy and lower computing power requirements.

[0005] The present application embodiment provides a method, including:

[0006] Collect monitoring data corresponding to the substation, meteorological data corresponding to the substation, and calendar information data, wherein the calendar information data is used to characterize the working hours corresponding to the substation and the non-working hours corresponding to the substation, the monitoring data is used to characterize the operating conditions of the substation transportation unit in the substation, and the meteorological data is used to characterize the environmental conditions of the substation;

[0007] Generate area data based on the monitoring data, the meteorological data and the calendar information data, wherein the area data is used to characterize the overall condition of the area;

[0008] At least one target prediction interval is generated according to the station area data through a charging capacity interval prediction model, wherein each target prediction interval is used to characterize the charging capacity interval prediction results corresponding to each different prediction confidence level.

[0009] Specifically, generating the area data based on the monitoring data, the meteorological data and the calendar information data includes:

[0010] Performing vectorization processing on the monitoring data, the meteorological data and the calendar information data respectively to obtain a monitoring data vector, a meteorological data vector and a calendar information data vector;

[0011] The monitoring data vector, the meteorological data vector and the calendar information data vector are subjected to vector splicing processing to obtain the station area data.

[0012] Specifically, the charging capacity interval prediction model includes a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer and a fifth hidden layer. The charging capacity interval prediction model is used to generate at least one target prediction interval according to the station area data, including:

[0013] Input the stage area data into the first hidden layer, and output a first two-dimensional tensor;

[0014] Input the first two-dimensional tensor into the second hidden layer, and output a second two-dimensional tensor;

[0015] Inputting the second two-dimensional tensor into the third hidden layer, and outputting a third two-dimensional tensor;

[0016] Inputting the third two-dimensional tensor into the fourth hidden layer, and outputting a fourth two-dimensional tensor;

[0017] The fourth two-dimensional tensor is input into the fifth hidden layer, and a fifth two-dimensional tensor is output, wherein the fifth two-dimensional tensor is used to characterize each target prediction interval.

[0018] Specifically, before collecting the monitoring data corresponding to the substation area, the meteorological data corresponding to the substation area, and the calendar information data, the method further includes:

[0019] Obtain the historical data of available charging capacity corresponding to the area;

[0020] Extracting a validation set and a test set from the available charging capacity historical data;

[0021] Based on the available charging capacity historical data, an initial prediction interval is established by a fuzzy kernel density estimation method;

[0022] Based on the initial prediction interval and the available charging capacity historical data, a first initial prediction interval is selected from the initial prediction interval by a ten-fold cross validation method, wherein the model performance estimation value corresponding to the first initial prediction interval is the highest;

[0023] Determining the first initial prediction interval and the confidence interval corresponding to the first initial prediction interval as training sets corresponding to each initial charging capacity interval prediction model;

[0024] Based on the training set and the available charging capacity historical data, respectively training each of the initial charging capacity interval prediction models to obtain each trained initial charging capacity interval prediction model;

[0025] Based on the validation set and the test set, the trained initial charging capacity interval prediction model with the best model performance among the trained initial charging capacity interval prediction models is determined as the charging capacity interval prediction model.

[0026] Specifically, the step of obtaining the historical data of available charging capacity corresponding to the station area includes:

[0027] Collect the charging load history data corresponding to the area and the terminal unit history load data corresponding to the area;

[0028] The available charging capacity historical data is calculated based on the charging load historical data, the station terminal unit historical load data and the station capacity information corresponding to the station.

[0029] Specifically, the training set includes a first training set and a second training set, and based on the training set and the available charging capacity historical data, the initial charging capacity interval prediction models are trained respectively to obtain the trained initial charging capacity interval prediction models, including:

[0030] Based on the first training set, determining initial model parameters corresponding to each of the initial charging capacity interval prediction models;

[0031] generating a recursive formula based on the available charging capacity historical data and the second training set;

[0032] Based on the recursive formula and the second training set, the initial model parameters are updated to obtain model parameters;

[0033] Based on the model parameters, the initial charging capacity interval prediction models are respectively determined as the trained initial charging capacity interval prediction models, and the initial charging capacity interval prediction models correspond one-to-one to the trained initial charging capacity interval prediction models.

[0034] Specifically, the method further includes:

[0035] By using a preset neural network computing stick, the charging capacity interval prediction model is deployed in the edge intelligent terminal in the form of a preset format file to reduce the computing power requirement and memory usage of the prediction process, wherein the edge intelligent terminal includes a preset mobile neural network structure.

[0036] The embodiments of the present application provide a method, device, terminal and storage medium for predicting charging capacity intervals in a substation, which can collect monitoring data corresponding to the substation, meteorological data corresponding to the substation and calendar information data, wherein the calendar information data is used to characterize the working time corresponding to the substation and the non-working time corresponding to the substation, the monitoring data is used to characterize the operating conditions of the substation transportation units in the substation, and the meteorological data is used to characterize the environmental conditions of the substation; then, based on the monitoring data, the meteorological data and the calendar information data, the substation data is generated, wherein the substation data is used to characterize the overall condition of the substation; finally, through a charging capacity interval prediction model, each target prediction interval for characterizing the charging capacity interval prediction results corresponding to each different prediction confidence is generated according to the substation data.

[0037] Compared with the prior art, the embodiments of the present application can achieve real-time rolling prediction of the available charging capacity range of the station area with less memory usage and lower computing power requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 It is a flow chart of a method for predicting charging capacity intervals in a substation area provided in an embodiment of the present application;

[0040] Figure 2 is a schematic diagram of a trapezoidal fuzzy set provided in an embodiment of the present application;

[0041] Figure 3 is a structural timing diagram of the training OS-ELM model provided in an embodiment of the present application;

[0042] Figure 4 This is a schematic diagram of rolling prediction provided by an embodiment of the present application;

[0043] Figure 5 is a schematic diagram of prediction results provided in an embodiment of the present application;

[0044] Figure 6 is a schematic diagram of a hardware device provided in an embodiment of the present application;

[0045] Figure 7 It is a schematic diagram of the network structure of the OS-ELM model provided in the embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0047] The embodiment of the present application provides a method for predicting a charging capacity interval of a station area.

[0048] Specifically, the device may be integrated into an electronic device, and the electronic device may be a terminal, a server, or other device.

[0049] In some embodiments, the server may also be implemented in the form of a terminal.

[0050] Among them, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), as well as big data and artificial intelligence platforms.

[0051] The terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0052] It should be noted that the serial numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0053] In this embodiment, a method for predicting charging capacity intervals in a substation is provided. The method can achieve real-time rolling prediction of available charging capacity intervals in a substation with less memory usage and lower computing power requirements.

[0054] like Figure 1 As shown, the specific process of the charging capacity interval prediction method of the substation can be as follows:

[0055] 110. Collect monitoring data corresponding to the substation, meteorological data corresponding to the substation, and calendar information data, wherein the calendar information data is used to characterize the working hours corresponding to the substation and the non-working hours corresponding to the substation, the monitoring data is used to characterize the operating status of the substation transportation unit in the substation, and the meteorological data is used to characterize the environmental status of the substation.

[0056] Substation is an important concept in the power system, usually refers to a smaller power supply unit, including transformers, distribution equipment, energy meters, etc., used to provide power supply to a specific area or building. The role of the substation is to convert the electric energy of the high-voltage transmission line into electric energy suitable for low-voltage distribution, and supply the electric energy to the user's terminal equipment. The following are some characteristics and functions of the substation:

[0057] Power conversion: The transformer in the substation can convert the high-voltage power transmitted by the transmission line into power suitable for low-voltage distribution to meet the needs of users;

[0058] Power distribution: The power distribution equipment in the substation can distribute power to different users or buildings to achieve the distribution and management of power supply;

[0059] Electricity metering: Electricity meters are usually installed in the substation area to measure the user's electricity consumption for electricity bill settlement;

[0060] Fault isolation: The power distribution equipment in the substation also includes a fault isolation device, which can quickly isolate the fault area when a fault occurs to ensure normal power supply to other users;

[0061] Remote monitoring: Modern substations are also equipped with remote monitoring systems that can monitor power load, equipment status and other information in real time to achieve remote management and control.

[0062] In some embodiments, the station may be a power supply unit of an electric vehicle.

[0063] 120. Generate substation area data based on the monitoring data, the meteorological data, and the calendar information data, wherein the substation area data is used to characterize the overall condition of the substation area.

[0064] In some embodiments, the generating of the station area data based on the monitoring data, the meteorological data and the calendar information data includes the following steps 121 to 122:

[0065] 121. Perform vectorization processing on the monitoring data, the meteorological data, and the calendar information data respectively to obtain a monitoring data vector, a meteorological data vector, and a calendar information data vector.

[0066] 122. Perform vector splicing processing on the monitoring data vector, the meteorological data vector and the calendar information data vector to obtain the station area data.

[0067] 130. Generate at least one target prediction interval according to the station area data through a charging capacity interval prediction model, wherein each target prediction interval is used to characterize a charging capacity interval prediction result corresponding to each different prediction confidence level.

[0068] In some embodiments, the charging capacity interval prediction model includes a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and a fifth hidden layer.

[0069] In some embodiments, the generating at least one target prediction interval according to the station area data by using the charging capacity interval prediction model includes steps 131 to 135 as shown below:

[0070] 131. Input the station area data into the first hidden layer, and output a first two-dimensional tensor.

[0071] 132. Input the first two-dimensional tensor into the second hidden layer, and output a second two-dimensional tensor.

[0072] 133. Input the second two-dimensional tensor to the third hidden layer, and output a third two-dimensional tensor.

[0073] 134. Input the third two-dimensional tensor to the fourth hidden layer, and output a fourth two-dimensional tensor.

[0074] 135. Input the fourth two-dimensional tensor into the fifth hidden layer, and output a fifth two-dimensional tensor, wherein the fifth two-dimensional tensor is used to characterize each target prediction interval.

[0075] In some embodiments, before collecting the monitoring data corresponding to the substation area, the meteorological data corresponding to the substation area, and the calendar information data, the method further includes steps A1 to A7 as shown below:

[0076] A1. Obtain the historical data of available charging capacity corresponding to the area.

[0077] A2. Extracting a verification set and a test set from the available charging capacity historical data.

[0078] A3. Based on the available charging capacity historical data, an initial prediction interval is established by using a fuzzy kernel density estimation method.

[0079] A4. Based on the initial prediction interval and the available charging capacity historical data, a first initial prediction interval is selected from the initial prediction intervals through a ten-fold cross validation method, wherein the model performance estimate corresponding to the first initial prediction interval is the highest.

[0080] A5. Determine the first initial prediction interval and the confidence interval corresponding to the first initial prediction interval as training sets corresponding to each initial charging capacity interval prediction model.

[0081] A6. Based on the training set and the available charging capacity historical data, each of the initial charging capacity interval prediction models is trained to obtain each trained initial charging capacity interval prediction model.

[0082] A7. Based on the validation set and the test set, the trained initial charging capacity interval prediction model with the best model performance among the trained initial charging capacity interval prediction models is determined as the charging capacity interval prediction model.

[0083] In some embodiments, the step of obtaining the available charging capacity history data corresponding to the station area includes the following steps B1 to B2:

[0084] B1. Collect the charging load history data corresponding to the substation and the terminal unit history load data corresponding to the substation.

[0085] B2. Calculate the available charging capacity historical data based on the charging load historical data, the station terminal unit historical load data and the station capacity information corresponding to the station.

[0086] In some embodiments, the training set includes a first training set and a second training set, and based on the training set and the available charging capacity historical data, the initial charging capacity interval prediction models are trained respectively to obtain the trained initial charging capacity interval prediction models, including the steps C1 to C4 as shown below:

[0087] C1. Based on the first training set, determine the initial model parameters corresponding to each of the initial charging capacity interval prediction models.

[0088] C2. Generate a recursive formula based on the available charging capacity historical data and the second training set.

[0089] C3. Based on the recursive formula and the second training set, update the initial model parameters to obtain model parameters.

[0090] C4. Based on the model parameters, each of the initial charging capacity interval prediction models is determined as the each trained initial charging capacity interval prediction model, and the initial charging capacity interval prediction model corresponds to the trained initial charging capacity interval prediction model one by one.

[0091] In some embodiments, the method further includes the following specific implementation contents:

[0092] By using a preset neural network computing stick, the charging capacity interval prediction model is deployed in the edge intelligent terminal in the form of a preset format file to reduce the computing power requirement and memory usage of the prediction process, wherein the edge intelligent terminal includes a preset mobile neural network structure.

[0093] In summary, compared with the prior art, this embodiment can achieve real-time rolling prediction of the available charging capacity range of the station area with less memory usage and lower computing power requirements.

[0094] The present application is further described below through a specific embodiment:

[0095] S1: Collect the charging load history data of the area and the TTU history load data of the area, that is, the charging load history data corresponding to the area and the terminal unit history load data corresponding to the area, and calculate the available charging capacity history data of the electric vehicles in the area based on the capacity information of the area, that is, calculate and obtain the aforementioned available charging capacity history data. At the same time, identify the outliers of the historical data, interpolate the missing values, and perform normalization processing to unify the dimensions, divide the available charging capacity history data, and obtain the training set, validation set, and test set.

[0096] Specifically, S1 includes the following specific steps:

[0097] (1) Collect historical operation data of the substation, including: historical charging load data of electric vehicles in the substation L c ,,TTU historical load data of the area L T And the distribution transformer rated capacity C T . Historical data of available charging capacity of electric vehicles in the area P c The calculation is as follows:

[0098] P c =0.8×C T -(L T -L c )

[0099] (2) Isolation Forests (IFS) is used to identify outliers in historical data. IFS can calculate the outlier score of each time series sample. The higher the score, the greater the possibility that the data is an outlier. The calculation process is as follows: input the historical data sample into the iTree tree, iterate along the corresponding conditional tree branch until the terminal leaf node, and record the path length to complete the calculation of the anomaly score.

[0100] (3) The missing data are interpolated using multi-stage interpolation with chain equations (MICE). The basic idea is to treat the missing variables as the dependent variables in the regression equation, and the non-missing data as the predictor variables, and interpolate the missing data using chain equations multiple times. After data preprocessing, data normalization is required to avoid dimensional differences affecting the model's fitting performance.

[0101] (4) Divide the historical available charging capacity data into N sub subintervals, each of which contains the same sample size Ns. 80% of the data is used as a training set, 10% of the data is used as a validation set, and 10% of the data is used as a test set.

[0102] S2. The probability density function (PDF) of each segmented sub-interval is calculated based on the fuzzy kernel density estimation method to fit the data distribution characteristics of the available charging capacity. The fuzzy interval modeling technology is used to select the optimal kernel bandwidth, and the flexible fine-tuning of the kernel density estimator is achieved through trapezoidal fuzzy sets to cope with the uncertainty brought to the charging behavior of car owners by related exogenous variables such as holidays and meteorological conditions. By adjusting the significance level parameters, the available charging capacity intervals (capacity lower bound / upper bound) at three confidence levels of 90% / 80% / 70% are constructed.

[0103] Specifically, step S2 includes:

[0104] (1) Fuzzy kernel density estimation method: The kernel density estimator is used to characterize the uncertainty of the available charging capacity of electric vehicles in the station area, and the probability density function (PDF) is used to fit the data distribution characteristics of the available charging capacity to construct a high-quality prediction interval. The flexibility interval of the available charging capacity is calculated by the kernel density estimation (KDE) estimator. Approximately, when N is given s Independent random power samples hour, It can be derived based on a general linear partial differential equation (PDE) with smooth properties. Specifically, it is shown in the following formula:

[0105]

[0106]

[0107] Where Δx is the empirical density of data in the initial conditions; It means that the linear differential operator is composed of two arbitrary positive functions α(x) and p(x); represents the difference operator, x is the function independent variable; X i is the i-th element of the estimator; in addition, boundary conditions must be considered to ensure The integral of is 1, as shown in the following formula:

[0108] L(·)=d(a(x)d(· / p(x)) / dx) / 2dx

[0109]

[0110] in, represents a differential operator if p(x) is a valid PDF for the dataset X and The solution in the above formula can be regarded as the flexibility kernel density estimate, and p(x) is the pilot PDF. When α(x)=p(x) λ When λ∈[0,1], the differential operator is adjustable. Thus, the estimator The general solution form is as follows:

[0111]

[0112]

[0113] x∈X,t>0

[0114]

[0115] y∈X,t>0 ·

[0116] Among them, in the above two linear partial differential equations, x and y correspond to the independent variable and the dependent variable respectively. for The adjoint operator of . is the kernel function. For the time series of the available charging capacity in the substation, its boundary conditions should ensure is an effective PDF, and the integral sum is 1, as shown in the following formula:

[0117]

[0118]

[0119] When the parameters x and y satisfy the above constraints, the general form of the kernel density estimator (DiE) can be given by:

[0120]

[0121] Among them, h DiE is the blur kernel κ DiE bandwidth (BW), and satisfies κ DiE (x,y;0)=δ(xy). is the fuzzy coefficient and assists the fuzzy kernel to fuzzify the initial density Δx at different rates, thereby extracting the PDF distribution characteristics of the available charging capacity time series and identifying the smoothing characteristics of the original time series within the confidence interval. If a(x)=p(x)=1, the above The partial differential equation is the Fourier heat conduction equation, which can be solved by Gaussian kernel density estimation. The analysis of a(x) and p(x) will be described in detail in the next section. In addition, selecting the optimal fuzzy kernel bandwidth can significantly improve the performance of the kernel density estimator, which is also the key to accurately modeling the PDF of the available charging capacity of electric vehicles in the area.

[0122] (2) Optimal selection strategy for fuzzy kernel bandwidth: DiE determines BW as the objective function with asymptotic minimum, where the embedded kernel density estimator The bandwidth of the pilot blur kernel τ DiE and the blur kernel κ DiE Make an estimate. Because E f [σ -1 (x)], L * (·) depends on a(x) and p(x). First, we need to estimate p(x). Then, by adjusting λ∈[0,1], a(x) can be obtained by the formula a(x)=p(x) λ The specific formula is as follows:

[0123]

[0124]

[0125]

[0126] in It means to find the expected value.

[0127] (3) Estimation of capacity interval (upper and lower bounds of available capacity): Based on the PDF obtained above, calculate the cumulative distribution function F of the response t(·) (x), and thus find the corresponding upper / lower quantiles as well as Finally, the prediction interval bandwidth of the available charging capacity of electric vehicles in the area based on confidence is constructed, as shown in the following formula:

[0128]

[0129]

[0130]

[0131] Among them, α l =α / 2,α u =1-α / 2, α is the significance level, the confidence interval is CL=100×(1-α)%, and Int. is the prediction interval bandwidth.

[0132] (4) Adaptive fine-tuning of capacity intervals based on trapezoidal fuzzy sets: In order to characterize the uncertainty of the available charging capacity in the station area by estimating the true PDF on continuous subintervals, DiE utilizes Figure 2 The trapezoidal fuzzy set (μ) is shown on the left axis. Based on the average value of the time series samples in each subinterval, the fuzzy set fine-tunes the flexibility of DiE to avoid boundary effects, as shown in the following formula:

[0133]

[0134] Among them, μ L , μ M , μ H They represent the low available charging capacity interval, the medium available charging capacity interval, and the high available charging capacity interval respectively. Characterizing the coverage of the entire sub-interval, if the average value of the sub-interval is near the boundary, the trapezoidal fuzzy set can effectively alleviate the problem. The boundary effect will affect the low available charging capacity interval And high available charging capacity range For the medium power range Boundary effects have little impact on it, DiE chooses λ=μ M (0<μ M ≤1) to ensure the smoothness of the PDF while preserving the timing characteristics of the available charge capacity.

[0135] S3. Determine the optimal prediction step size, the optimal rolling window, and the number of hidden layer nodes of OS-ELM through ten-fold cross validation, select the optimal sub-interval and the corresponding upper and lower bounds of the confidence interval as the training set of the OS-ELM model, and the OS-ELM model is the charging capacity interval prediction model.

[0136] Specifically, step S3 includes:

[0137] (1) The extreme learning machine (ELM) algorithm is different from traditional machine learning algorithms, which require iterative gradient training. Instead, it randomly selects input weights and biases, and then determines the output weights through simple matrix calculations. It avoids many limitations of traditional neural network training algorithms, such as slow learning process, local minima, overfitting, large amount of computation, and heavy training burden. For any infinitely differentiable activation function, the ELM with hidden layer neurons can accurately learn different samples with zero error. It has faster learning speed and excellent generalization ability than the backpropagation algorithm. As a lightweight algorithm, it can run efficiently on the edge with limited computing power.

[0138] As a batch-based algorithm, ELM needs to train all training samples at once during the training phase and then verify the test set. OS-ELM is an online learning mode that can update model parameters as new training samples are added, that is, update network parameters online. It has the advantages of high algorithm efficiency and strong adaptability of ELM, and can realize real-time rolling prediction on the edge side. OS-ELM is divided into two processes. The first process is to initialize model parameters using the ELM algorithm with a small number of training samples; the second process starts online learning. Every time a new training sample arrives, the new network parameters are solved according to the recursive formula, thereby realizing the online training process.

[0139] (2) To achieve rolling prediction, each rolling window N lag The output of the OS-ELM model is g i Including the upper quantile and lower quantile of the available capacity of electric vehicles at future time nodes, such as Figure 3 As shown. The prediction step length is N = N sub -N lag 10-fold cross-validation is a commonly used model evaluation method to estimate the performance of the model on unknown data. The steps are as follows: randomly divide the original data into N sub Subsets (i.e. folds) of equal size. Use N of them in turn sub -1 fold as the training set, and the remaining fold as the test set, train the model and evaluate the performance on the test set. Repeat step 2 N sub times, each time selecting a different fold as the test set. Calculate the prediction error for each test. Calculate N subThe average value of the evaluation index is used as the estimated value of the model performance. The 10-fold cross validation can avoid the problem of overfitting or underfitting of the model on a single validation set. At the same time, it can ensure that the model has stronger generalization performance. The best sub-interval is selected as the optimal training set of the model for training the final OS-ELM model.

[0140] S4. Use the optimal sub-interval training set data to perform offline parallel training on the PC side for N OS-ELM models, where the input is the TTU monitoring data of the substation with a sampling frequency of 15 minutes (the latest 96 sampling points, a total of 24 hours of historical data), the corresponding meteorological information and calendar information, and the output is a 16×6 two-dimensional tensor, that is, the upper / lower bounds of the available charging capacity at three confidence levels of 90% / 80% / 70% in the next 4 hours (16 sampling points) of the substation. Through iterative training of historical data, calculate the optimal output weight vector and related parameters of each OS-ELM model. Meteorological data also includes 96 sampling points (temperature, humidity, precipitation, air pressure), that is, a 96×4 array matrix (which can also be understood as a multidimensional vector). Calendar information is divided into working hours (9:00-17:00 am is defined as working hours) and non-working hours. Working hours are represented by 1 and non-working hours are represented by 0 (all weekends and holidays are represented by 0). Therefore, calendar information is a 96×1 vector (vector elements can only take values ​​0 or 1). Therefore, the input of the model is the concatenation of the monitoring data, calendar information, and meteorological information of the most recent 96 sampling points, which is a 96×6 array matrix in total. Input the first hidden layer to obtain a 96×32 two-dimensional tensor, input the second hidden layer to obtain a 32×64 two-dimensional tensor, input the third hidden layer to obtain a 64×16 two-dimensional tensor, input the fourth hidden layer to obtain a 16×32 two-dimensional tensor, and input the fifth hidden layer to finally obtain a 16×6 two-dimensional tensor. That is, corresponding to the next 4 hours (16 sampling points) of the substation, every 6 points correspond to the upper / lower bounds of the available charging capacity under three confidence levels of 90% / 80% / 70%, thereby obtaining three available charging capacity intervals.

[0141] Specifically, S4 includes:

[0142] (1) Random feature mapping: First, define the training sample as The ELM with N hidden nodes can be expressed as x i represents the input feature column vector, g i represents the target output; Figure 4 As shown, β is the output weight matrix, H is the hidden layer output matrix, and G is the target matrix. The output value of each neural node in the hidden layer is given a certain weight, and then the output value of each neural node is superimposed to obtain the output value. The matrix form is expressed as:

[0143]

[0144] G=Hβ

[0145]

[0146] In the above formula, β represents the connection weight between the Lth neural node in the hidden layer and the output layer, that is, the output weight, and G represents the output value during the algorithm training process. There are three unknown parameters in the ELM model corresponding to the OS-ELM model, namely, the connection weight W of the hidden layer i , bias vector b i And the output weight β. The learning process of the ELM model is the process of continuously updating parameters according to the training samples. After completing the forward transfer, the ELM model will solve these three values ​​according to its own mechanism.

[0147] (2) Linear parameter solution: The second stage of ELM model training is mainly to solve the output weight β. The purpose of training is to reduce the error. Therefore, the purpose of adjusting the parameter β is to minimize the error between the model output value and the expected output value. Here, we calculate the square difference between Hβ and the expected output value G as the error function, and the objective function of training is the error function. The solution β that minimizes the objective function value is the optimal parameter selection for the model. The objective function is as follows:

[0148]

[0149] min||Hβ-T|| 2 ,β∈R L×m

[0150] Where T is the expected output matrix, are the corresponding elements in the matrix.

[0151] According to the formula and linear algebra knowledge, the optimal solution of β is shown in the following formula, where H' is the Moore-Penrose generalized inverse matrix of matrix H:

[0152]

[0153] Therefore, the optimal parameter solution process of the ELM model can be transformed into the generalized inverse matrix of H. There are many ways to solve the generalized inverse matrix, such as orthogonal projection method, orthogonalization method, iteration method and singular value decomposition (SVD) method. T When H is a reversible matrix, it can be solved in the following way:

[0154] H'=(HTH)-1HT

[0155] In actual use, there are often H TWhen H is a singular matrix, the orthogonal projection method is not applicable to all cases. In programming languages ​​such as Python, the function of solving the generalized inverse matrix Moore-Penrose can be called to obtain the optimal output weight β, thus completing the ELM training process.

[0156] (3) The algorithm flow of OS-ELM is as follows: First, the initial weight β0 is calculated by the above formula

[0157]

[0158] Among them, H0 and T0 represent the initial state, When new training samples enter the OS-ELM model, for example, N1 samples are added to the model, the formula for minimizing the objective function of the extreme learning machine is as follows:

[0159]

[0160] By solving the generalized inverse matrix, we can get β (1) as follows:

[0161]

[0162]

[0163] For the online learning process, the new parameters need to be represented by historical parameters and newly input training samples, that is, β (1) It is expressed as a function of β0, K1, H1 and T1, where K1 can be expressed by the following formula:

[0164]

[0165] Further reasoning leads to:

[0166]

[0167] It can be inferred from this that β (1) :

[0168]

[0169] According to the above formula, the recursive formula of the online learning model can be derived:

[0170]

[0171] and It can be solved according to Woodbury formula:

[0172]

[0173] set up Then the formula can be expressed as:

[0174]

[0175]

[0176] If the training samples are updated one by one into the ELM model, the formula can be simplified to the following form.

[0177]

[0178]

[0179] The above is the principle of the OS-ELM algorithm. First, the model parameters can be preliminarily determined based on a small number of training samples. Then, two recursive formulas can be used to adjust and update the model parameters based on the newly added training samples to achieve the purpose of online learning and rolling prediction. Figure 4 shown.

[0180] S5. Conduct data experiments on the validation set, use sharpness, reliability, and comprehensive scores as evaluation indicators to evaluate the performance of each sub-model OS-ELM, select the best OS-ELM sub-model, package the weight data and OS-ELM program into a lightweight dedicated file BSON format for use by the Huawei Ascend 310Atlas 200DK edge intelligent platform for on-site prediction, and deploy the MobileNet network structure and use Movidius computing sticks and other conversion models on the edge intelligent terminal Atlas in advance to ensure low computing power requirements and less memory usage during the prediction process. OS-ELM monitors the real-time TTU data of the substation area, the corresponding weather forecast information, and calendar information; outputs the upper / lower bounds of the available charging capacity at three confidence levels of 90% / 80% / 70% for the next 4 hours (16 sampling points) in the substation area, and updates the iterative model online to ensure the model's adaptability and generalization capabilities, and ultimately achieve rolling interval prediction of the available charging capacity of electric vehicles. Specifically:

[0181] (1) In the aforementioned step S4, OS-ELM1 to OS-ELM N The best training data set has been used for parallel training. Input the validation set data into the trained OS-ELM1 to OS-ELM N The model generates the available charging capacity intervals PI1 to PI N , based on the sharpness and reliability indicators, the accuracy of the prediction interval PI is evaluated, so as to select the best OS-ELM model. The details are as follows:

[0182] 1) Reliability: The deviation of the PI prediction interval coverage probability (PICP) is defined as the average coverage error (ACE). The deviation of PICP from the nominal interval coverage probability (PINC) should be positive, and the closer it is to 0, the more reliable the prediction interval PI is and the more data distribution characteristics it can retain. The specific definition is as follows:

[0183] 0≤ACE=PICP-PINC<ε.

[0184]

[0185]

[0186] in is the indicator function, N test is the number of test set samples, and are the upper and lower bounds of the available charging capacity time series data associated with the prediction target. PICP is used to characterize the probability density distribution of the available charging capacity of electric vehicles in the future area, where the prediction target g i , will be surrounded by the prediction interval.

[0187] 2) Sharpness: In order to extract feature information from the PI interval, the normalized average width (PINAW) in the following formula should take a smaller value to produce a PI with higher sharpness.

[0188]

[0189] Where R is the target range used to normalize the PI averaging bandwidth.

[0190] 3) Comprehensive score: In order to evaluate the overall performance of the PI interval construction method, the overall score in the following formula is used in the evaluation process, which takes into account both reliability and sharpness. The sharper the PI interval, the smaller the value |SC|.

[0191]

[0192] Since a very sharp prediction interval PI may violate the reliability standard, the comprehensive score indicator Sc can reasonably reflect the accuracy of the prediction interval in characterizing the uncertainty of the available capacity of electric vehicles in the station area.

[0193] (2) Obtain the best OS-ELM through the above process optModel, convert the trained model into an offline model format (BSON file) that can run on the Ascend AI processor. Use the Model Compiler (ATC) tool provided by the Ascend SDK to convert the trained model into a BSON model file used on the Ascend processor. Using the development environment provided by the Ascend SDK, use C++ or Python to write and optimize applications, and interact with the Ascend 310Atlas 200DK through the API interface. Copy the optimized application and the converted BSON model file to the Atlas200DK edge computing module at the same time. The Movidius neural network computing stick is a hardware accelerator developed by Intel to accelerate deep learning reasoning tasks on edge computing devices. It can be connected to embedded devices through a USB interface. Perform deep learning reasoning tasks with low power consumption. Using the Movidius neural network computing stick, the neural network model can be deployed to the edge device 310Atlas 200DK without relying on cloud computing resources. Real-time deep learning reasoning tasks can be performed locally on the device to improve computing efficiency and response speed.

[0194] Configure the communication parameters for connecting to the low-voltage cabinet on the substation intelligent fusion terminal (TTU), including IP address, port number, access rights, etc. Determine the protocol for communicating with the low-voltage control cabinet, such as Modbus, DNP3, etc. Ensure that the substation intelligent fusion terminal and the low-voltage control cabinet use the same communication protocol. The substation intelligent fusion terminal can establish a connection with the low-voltage cabinet through the 4G / 5G network. The connection can be achieved using the API or configuration tool provided by the communication module. Once the connection is established, the substation intelligent fusion terminal can transmit the real-time load information of the substation to the Atlas 200Dk edge smart chip of the low-voltage cabinet. Low-voltage power line high-speed carrier communication technology (HLPC) is mature and stable, and is suitable for signal communication of low-voltage power line networks, including two-way data transmission, remote control and positioning services. In this case, the HLPC communication module can be used to connect the low-voltage cabinet and the smart charging pile to the communication network and realize data transmission and exchange. The Atlas 200Dk module predicts the available charging capacity range at future time nodes based on the real-time collected TTU load data and charging pile charging load data. The Atlas 200Dk prediction output format is as follows: Figure 5 As shown, the structure of the hardware device is as follows Figure 6 OS-ELM iteratively updates the network model parameters based on real-time data and introduces the MobileNet deep separable convolution operation, which significantly reduces the number of parameters and computational complexity in the network while maintaining strong feature expression capabilities and achieving efficient edge training.

[0195] From the above, it can be seen that the embodiments of the present application can collect monitoring data corresponding to the substation, meteorological data corresponding to the substation, and calendar information data, wherein the calendar information data is used to characterize the working time corresponding to the substation and the non-working time corresponding to the substation, the monitoring data is used to characterize the operating conditions of the substation transportation units in the substation, and the meteorological data is used to characterize the environmental conditions of the substation; then, based on the monitoring data, the meteorological data and the calendar information data, the substation data is generated, wherein the substation data is used to characterize the overall condition of the substation; finally, through the charging capacity interval prediction model, each target prediction interval for characterizing the charging capacity interval prediction results corresponding to each different prediction confidence is generated according to the substation data.

[0196] In summary, this solution provides a method for predicting the charging capacity interval of a substation. Compared with the existing technology, it can realize real-time rolling prediction of the available charging capacity interval of the substation with less memory occupancy and lower computing power requirements.

[0197] A person skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a storage medium and loaded and executed by a processor.

[0198] To this end, the embodiment of the present application may also provide a storage medium in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any of the methods for predicting the charging capacity interval of a station area provided in the embodiment of the present application. For example, the instructions may execute the following steps:

[0199] Collect monitoring data corresponding to the substation, meteorological data corresponding to the substation, and calendar information data, wherein the calendar information data is used to characterize the working hours corresponding to the substation and the non-working hours corresponding to the substation, the monitoring data is used to characterize the operating conditions of the substation transportation unit in the substation, and the meteorological data is used to characterize the environmental conditions of the substation;

[0200] Generate area data based on the monitoring data, the meteorological data and the calendar information data, wherein the area data is used to characterize the overall condition of the area;

[0201] At least one target prediction interval is generated according to the station area data through a charging capacity interval prediction model, wherein each target prediction interval is used to characterize the charging capacity interval prediction results corresponding to each different prediction confidence level.

[0202] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0203] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, the computer instructions are stored in a storage medium. A processor of a computer device reads the computer instructions from the storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided in various optional implementations provided in the above embodiments.

[0204] Since the instructions stored in the storage medium can execute the steps in any of the methods for predicting the charging capacity interval of a substation provided in the embodiments of the present application, the beneficial effects that can be achieved by any of the methods for predicting the charging capacity interval of a substation provided in the embodiments of the present application can be achieved. For details, please refer to the previous embodiments and will not be repeated here.

[0205] The above is a detailed introduction to a method for predicting the charging capacity interval of a substation provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for predicting charging capacity intervals in a substation area. It is characterized in that include: Collect monitoring data corresponding to the substation, meteorological data corresponding to the substation, and calendar information data, wherein the calendar information data is used to characterize the working hours corresponding to the substation and the non-working hours corresponding to the substation, the monitoring data is used to characterize the operating conditions of the substation transportation unit in the substation, and the meteorological data is used to characterize the environmental conditions of the substation; Generate area data based on the monitoring data, the meteorological data and the calendar information data, wherein the area data is used to characterize the overall condition of the area; At least one target prediction interval is generated according to the station area data through a charging capacity interval prediction model, wherein each target prediction interval is used to characterize the charging capacity interval prediction results corresponding to each different prediction confidence level.

2. A method for predicting charging capacity intervals in a substation as claimed in claim 1, It is characterized in that The generating of the station area data based on the monitoring data, the meteorological data and the calendar information data comprises: Performing vectorization processing on the monitoring data, the meteorological data and the calendar information data respectively to obtain a monitoring data vector, a meteorological data vector and a calendar information data vector; The monitoring data vector, the meteorological data vector and the calendar information data vector are subjected to vector splicing processing to obtain the station area data.

3. A method for predicting charging capacity intervals in a substation as claimed in claim 1, It is characterized in that The charging capacity interval prediction model includes a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer and a fifth hidden layer. The charging capacity interval prediction model is used to generate at least one target prediction interval according to the station area data, including: Input the stage area data into the first hidden layer, and output a first two-dimensional tensor; Inputting the first two-dimensional tensor into the second hidden layer, and outputting a second two-dimensional tensor; Inputting the second two-dimensional tensor into the third hidden layer, and outputting a third two-dimensional tensor; Inputting the third two-dimensional tensor into the fourth hidden layer, and outputting a fourth two-dimensional tensor; The fourth two-dimensional tensor is input into the fifth hidden layer, and a fifth two-dimensional tensor is output, wherein the fifth two-dimensional tensor is used to characterize each target prediction interval.

4. A method for predicting charging capacity intervals in a substation as claimed in claim 1, It is characterized in that Before collecting the monitoring data corresponding to the substation area, the meteorological data corresponding to the substation area, and the calendar information data, the method further includes: Obtain the historical data of available charging capacity corresponding to the area; Extracting a validation set and a test set from the available charging capacity historical data; Based on the available charging capacity historical data, an initial prediction interval is established by a fuzzy kernel density estimation method; Based on the initial prediction interval and the available charging capacity historical data, a first initial prediction interval is selected from the initial prediction interval by a ten-fold cross validation method, wherein the model performance estimation value corresponding to the first initial prediction interval is the highest; Determining the first initial prediction interval and the confidence interval corresponding to the first initial prediction interval as training sets corresponding to each initial charging capacity interval prediction model; Based on the training set and the available charging capacity historical data, respectively training each of the initial charging capacity interval prediction models to obtain each trained initial charging capacity interval prediction model; Based on the validation set and the test set, the trained initial charging capacity interval prediction model with the best model performance among the trained initial charging capacity interval prediction models is determined as the charging capacity interval prediction model.

5. A method for predicting charging capacity intervals in a substation as claimed in claim 4, It is characterized in that The obtaining of the historical data of available charging capacity corresponding to the station area includes: Collect the charging load history data corresponding to the area and the terminal unit history load data corresponding to the area; The available charging capacity historical data is calculated based on the charging load historical data, the station terminal unit historical load data and the station capacity information corresponding to the station.

6. A method for predicting charging capacity intervals in a substation as claimed in claim 4, It is characterized in that The training set includes a first training set and a second training set. Based on the training set and the available charging capacity historical data, each initial charging capacity interval prediction model is trained to obtain each trained initial charging capacity interval prediction model, including: Based on the first training set, determining initial model parameters corresponding to each of the initial charging capacity interval prediction models; generating a recursive formula based on the available charging capacity historical data and the second training set; Based on the recursive formula and the second training set, the initial model parameters are updated to obtain model parameters; Based on the model parameters, the initial charging capacity interval prediction models are respectively determined as the trained initial charging capacity interval prediction models, and the initial charging capacity interval prediction models correspond one-to-one to the trained initial charging capacity interval prediction models.

7. A method for predicting charging capacity intervals in a substation as claimed in claim 1, It is characterized in that The method further comprises: By using a preset neural network computing stick, the charging capacity interval prediction model is deployed in the edge intelligent terminal in the form of a preset format file to reduce the computing power requirement and memory usage of the prediction process, wherein the edge intelligent terminal includes a preset mobile neural network structure.

Citation Information

Patent Citations

  • Photovoltaic power interval prediction method based on improved extreme learning machine and quantile regression

    CN112101669A

  • Medium and long term power load prediction method

    CN114091782A

  • Electric vehicle charging demand prediction method and device

    CN114331542A

  • Area available capacity prediction method and system, electronic equipment and storage medium

    CN117541303A

  • Probabilistic wind speed forecasting method and system based on multi-scale information

    US20230169230A1

Cited By

  • Ammeter abnormity identification method and system based on MobileNetV3

    CN119359695A