CN115526398A - Attention mechanism-based charging load prediction method for charging station - Google Patents
Attention mechanism-based charging load prediction method for charging station Download PDFInfo
- Publication number
- CN115526398A CN115526398A CN202211188648.7A CN202211188648A CN115526398A CN 115526398 A CN115526398 A CN 115526398A CN 202211188648 A CN202211188648 A CN 202211188648A CN 115526398 A CN115526398 A CN 115526398A
- Authority
- CN
- China
- Prior art keywords
- data
- charging
- prediction
- module
- time period
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
- H02J3/003—Load forecast, e.g. methods or systems for forecasting future load demand
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
- H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Theoretical Computer Science (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- General Physics & Mathematics (AREA)
- Educational Administration (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- General Business, Economics & Management (AREA)
- Operations Research (AREA)
- Marketing (AREA)
- Game Theory and Decision Science (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Power Engineering (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The application discloses a charging station charging load prediction method based on an attention mechanism, which comprises a prediction model, wherein the prediction model comprises a characteristic embedding module, a historical data coding module and a fusion prediction module; obtaining characteristic data of a charging station time period t and a charging load value y corresponding to the time period t from historical data t And the prediction model calculates a prediction charging load value, and obtains a convergent prediction model through historical data training so as to predict the charging load in a future period. According to the method, a novel prediction method is adopted to give the charging load according to the historical charging load data of the charging stationThe charging load change curves of the charging stations in the future time period are predicted, data support is better provided for construction planning and daily operation of the charging stations, and further quantitative analysis basis is provided for load prediction of the power distribution network.
Description
Technical Field
The application relates to a charging load prediction method of a charging station based on an attention mechanism, which comprises the technologies of general data processing, data prediction or optimization and the like, and belongs to the technical field of data mining, utilization and prediction.
Background
Because the charging load has obvious space-time characteristics, the charging load of the electric vehicle has completely different performances in different time and space, and the charging load curves of different charging stations are different, although a traditional classical autoregressive time series prediction model such as AR, MA, ARMA, ARIMA and the like has the characteristics of strong pertinence and good robustness, the learning freedom degree of the model is low, the introduction support degree of covariates is not high, but the data distribution difference between different items is large, and the method cannot accurately capture the data distribution difference, so the generalization performance of the method is poor. Compared with an autoregressive model, the shallow machine learning method based on the feature engineering, such as a regression analysis method, a gradient lifting decision tree and a support vector machine, has improved freedom, but depends on complex feature engineering and business experience, has higher extrapolation complexity and lower preparation for predicting missing data.
The prediction method based on deep learning gradually becomes a research hotspot in the field of data prediction due to the characteristics of high learning freedom and low dependence on artificial characteristic engineering. Such as the autoregressive recurrent neural network deep proposed by David Salinas et al, the method combines autoregressive and recurrent neural networks, introduces covariates, supports the prediction of missing items through similar expressions among similar items, is also subject to the defects of the recurrent neural network, and is difficult to capture long-term period and season information. In recent years, attention mechanism has been highlighted in capturing long-term dependency, so researchers have fused attention mechanism into deep learning models to further improve the accuracy of data prediction. However, in the prediction problem in the real application scene, a general method may not obtain a good prediction result, and in order to solve the prediction problem in the actual complex scene, a plurality of methods are often combined for prediction.
Disclosure of Invention
The purpose of the application is to predict the charging load change curve of each charging station in a given future time period by adopting a novel prediction method according to historical charging load data of the charging stations, better provide data support for construction planning and daily operation of the charging stations and further provide quantitative analysis basis for load prediction of a power distribution network.
In order to achieve the above purpose, the technical scheme of the application is to provide a charging load prediction method for a charging station based on an attention mechanism, which is characterized by comprising a prediction model, wherein the prediction model comprises a feature embedding module, a historical data coding module and a fusion prediction module;
the prediction model training steps are as follows:
acquiring characteristic data of a charging station time period t and a charging load value y corresponding to the time period t t All feature data constituting a feature vector r t (ii) a Extracting a feature vector r t Middle time and weather characteristic phaseThe off fraction is denoted g t ；
Acquiring historical data of a charging station in a period t of the past s days, including
And
respectively representing the date, time and weather characteristics of a time period t in the past s days and the load value in the corresponding time period;
the input vector of the historical data encoding module is u t And g t ；
The historical data encoding module is provided with a network basic module FA which is sequentially provided with a linear layer, a BatchNorm layer, a Tanh layer and a Dropout layer from input to output;
inputting the vector X into a network basic module FA to obtain FA (X);
historical data encoding module pair vector u t The process of (3) is a residual concatenation process:
K t ＝FA(u t +FA(u t ))，
for g t The same treatment yields: q t ＝FA(g t +FA(g t ))；
K t And Q t Normalization processing is carried out after multiplication to obtain feature data pair prediction y corresponding to time period t in historical data t Influence factor alpha of t ：
α t ＝softmax(K t ×Q t )
By alpha t To p is p t Weighting and calculating to obtain the output vector V of the historical data coding module t ：
The input vector of the fusion prediction module is O t ，O t ＝[r t ，V t ]；
The output of the fusion prediction module is:
wherein, W p And b p Respectively representing a weight matrix and an offset, which are updated weights after learning and training are needed in a prediction model;
a mean square error loss function LMSE is set,
performing error back propagation and updating the prediction model based on a random gradient descent algorithm and a momentum minimization loss algorithm until the prediction model converges to obtain a converged prediction model;
the prediction steps of the prediction model are as follows:
recording the time period to be predicted as t ', acquiring characteristic data of the time period t', wherein the characteristic data related to weather is acquired through weather forecast, and substituting the characteristic data into the converged prediction model for calculation to obtain the characteristic data
I.e. the predicted charging load value for time period t'.
Wherein the feature data comprises x t ＝[x t0 ，x t1 ，...，x td-1 ]And h t ＝[h t0 ，k t1 ，...，h tk-1 ]，x t A continuous feature vector h composed of d continuous feature data corresponding to the time interval t t A category type feature vector composed of k category type feature data corresponding to the time period t; class-type feature vector h t Converting input feature embedding module into continuous feature vector h' t Continuous type feature vector x t And continuous type feature vector h' t Combining to obtain the characteristic vector r t ＝[h′ t |x t ]。
Specifically, the feature embedding module processes the input class-type feature vector as follows: one-hot coding is carried out on each feature, and then a learnable weight matrix is multiplied to convert the feature into a continuous feature vector with dimension r, wherein r is a hyper-parameter of the model and is determined through priori knowledge and manual debugging.
Specifically, the input-output relation of the linear layer is as follows:
f(X)＝XW T +b，
x represents an input vector of the layer, W and b represent a weight matrix and an offset respectively, and the weight matrix and the offset are updated weights after learning and training are needed in a prediction model;
the input and output relational expression of the BatchNorm layer is as follows:
mu represents the expectation of an input vector X, sigma is a standard deviation, epsilon is a positive number close to 0, gamma and beta respectively represent affine transformation parameter vectors and are updated weights after learning and training in a prediction model;
the input and output relational expression of the Tanh layer is as follows:
preferably, the feature data input to the prediction model is normalized by the following formula:
x' represents raw data, μ represents a mean value of the raw data, S represents a standard deviation of the raw data, and x is data after normalization processing.
Preferably, the charging load value is a charging load value after a logarithmic processing, and the logarithmic processing formula is:
y＝log(y′+1)
where y' represents the original data and y represents the data after the logarithmic processing.
Specifically, the softmax function formula is as follows:
preferably, the characteristic data at least includes charging station information, location information to which the charging station belongs, weather information, and time information. Preferably, the weather information includes a current time temperature and a current time humidity. Further, the time information includes the hour and the minute of the day where the current time is located.
Drawings
FIG. 1 is a flow diagram of a feature embedding process provided in an embodiment;
fig. 2 is a site charging load curve before the logarithm processing provided in the embodiment;
fig. 3 is a log-processed site charging load curve provided in the embodiment;
FIG. 4 is a schematic diagram of the overall structure of a prediction model provided in the embodiment;
fig. 5 is a comparison of site a charging load prediction results provided in the example;
fig. 6 is a comparison of site B charging load prediction results provided in the embodiments;
fig. 7 is a comparison of the site C charging load prediction results provided in the example;
fig. 8 is a comparison of the site D charging load prediction results provided in the example.
Detailed Description
In order to make the present application more comprehensible, preferred embodiments accompanied with figures are described in detail below.
Examples
The embodiment provides a charging load prediction method for a charging station based on an attention mechanism, which is used for predicting the charging load of the electric vehicle station aiming at a time period t, wherein the time period t can be selectively divided into minute levels, and the time division scale can be expanded based on practical application.
1. A model training stage:
the training stage of the model is mainly divided into three parts, 1, data acquisition and feature extraction; 2. preprocessing data; 3. and (5) training a model.
1. Data collection and feature extraction
In the method for predicting the load of the electric vehicle charging station, an operator of the charging equipment pushes data through the data transmission protocol in table 1, and my party receives the data and stores the data in the database.
TABLE 1 detail data of interface charging process of charging equipment
| Parameter name | Definition of | Must fill in | Type of parameter | Remarks for note |
| Charging order number | StartChargeSeq | Is that | Character string | Docking platform system order numbering |
| Charging device interface coding | ConnectorID | Is that | Character string | Unique muzzle number under platform |
| Vehicle identification code | Vin | Whether or not | Character string | Vehicle identification number or frame number |
| Phase A current | CurrentA | Is that | Floating point type | Unit: a, default: 0 with DC (output) |
| Phase voltage of A | VoltageA | Is that | Floating point type | Unit: v, default: 0 with DC (output) |
| Percentage of remaining battery capacity | SOC | Is that | Floating point type | Defaults are as follows: 0 |
| Time to start charging | StartTime | Is that | Character type | The format "yyyy-MM-dd HH: mm: ss' |
| This time of sampling | EndTime | Is that | Character type | The format "yyyy-MM-dd HH: mm: ss' |
| Cumulative charge amount | TotalElect | Is that | Floating point type | Unit: degree, 2 bits after decimal point |
| Accumulated electricity charge | ElecMoney | Is that | Floating point type | Unit: 2 bits after the element, decimal point |
| Accumulated service charge | SeviceMoney | Is that | Floating point type | Unit: 2 bits after the element, decimal point |
| Accumulated total sum | TotalMoney | Is that | Floating point type | Unit: 2 bits after the element, decimal point |
The power P (the sum of the powers taken by a station to the power system at a certain time is called a charging load of the station) can be obtained by multiplying the phase current a by the phase voltage a
P＝CurrentA×VoltageA (1)
And aggregating according to minutes, setting that k vehicles are charged within one minute of a certain station, and respectively pushing d to the charging behaviors of the k vehicles within the minute by an operator 1 、d 2 ...d k The charging load of the station in this minute is calculated according to the following formula (2):
meanwhile, the popularity of the charging station is related to the charging station and the geographical position of the charging station, and the charging behavior of the electric vehicle is influenced by weather and time, so that characteristic data coexistence tables such as charging station information, location information to which the charging station belongs, weather information, time information and the like are extracted from the database.
TABLE 2 characterization of the model
| Feature(s) | Feature classification | Description of the preferred embodiment |
| frequency | Continuous type | Monthly charging frequency at a site |
| conn_num | Continuous type | Number of charging interfaces of station |
| dc_conn_num | (Continuous)Model (III) | Number of site DC interfaces |
| ac_conn_num | Continuous type | Number of site AC interfaces |
| equipment_amount | Continuous type | Number of site devices |
| ac_equipment_amount | Continuous type | Number of site direct current devices |
| dc_equipment_amount | Continuous type | Number of station communication devices |
| st_equipment_power | Continuous type | Total power of site |
| dc_equipment_power | Continuous type | Total power of site DC equipment |
| single_power | Continuous type | Station single pile power |
| ac_dedicated | Continuous type | Line of siteNumber of communication devices dedicated to administrative district |
| dc_dedicated | Continuous type | Number of direct current devices dedicated to administrative district where site is located |
| ac_public | Continuous type | Number of public communication equipment in administrative district where site is located |
| dc_public | Continuous type | Number of common direct current devices in administrative district where site is located |
| density | Continuous type | Site distribution density of administrative district in which site is located |
| equipment_power | Continuous type | Charging equipment total power of administrative district where station is located |
| temperature | Continuous type | Temperature at present |
| humidity | Continuous type | Humidity of current time |
| power | Continuous type | SiteCharging load at present |
| district | Class type | Administrative district |
| hour | Class type | Hour(s) |
| minute | Class type | Minute (min) |
| station_type | Class type | Site type |
| park_fee_type | Class type | Type of stop parking charge |
| construction | Class type | Site construction site |
| demonstration | Class type | Charging or not demonstration station |
| city_area | Class type | The ring area to which the station belongs |
| equipment_structure | Class type | Station equipment composition structure |
| connector_structure | Class type | Station muzzle structure |
| scale | Class type | Size of operator to which site belongs |
The method specifically comprises four types of characteristic data, namely charging station information, location information to which the charging station belongs, weather information and time information. The charging station comprises a station, a station charging interface, a station direct current interface, a station alternating current interface, a station device, a station direct current device, a station alternating current device, a station total power, a charging demonstration station, a station direct current device total power, a station single pile power, a station type, a station parking charging type, a station construction site, a station device composition structure, a station gun mouth structure and a station affiliated operator scale, wherein the station previous month charging frequency, the station charging interface quantity, the station direct current interface quantity, the station device quantity, the station direct current device quantity, the station parking charging type, the station construction site, the station device composition structure, the station gun mouth structure and the station affiliated operator scale belong to charging station information; the number of the special alternating current equipment for the administrative district where the site is located, the number of the special direct current equipment for the administrative district where the site is located, the number of the public alternating current equipment for the administrative district where the site is located, the number of the public direct current equipment for the administrative district where the site is located, the site distribution density of the administrative district where the site is located, the total power of the charging equipment for the administrative district where the site is located, the administrative district, and the district to which the site belongs belong to the regional information of the charging site; the current time temperature and the current time humidity belong to weather information; the hour and the minute of the time period belong to the time information.
2. Data pre-processing
It can be seen from table 2 that the data set used in the method of the present invention mainly includes a class type variable and a continuous type variable, and for class type features, we convert the class variable into the continuous type variable by an Embedding manner (hereinafter referred to as Embedding), first perform one-hot encoding on each feature, and then convert into a continuous feature vector with dimension r by multiplying a learnable weight matrix, where r is a hyper-parameter of a model, and is determined by priori knowledge and manual debugging, and a feature Embedding processing flow is shown in fig. 1.
Because the random gradient descent algorithm is adopted for optimization in the model training process, under the condition that the fluctuation range of input data set data is large, the phenomena of gradient explosion and gradient disappearance can occur, so that the model cannot be converged. In order to prevent the phenomenon from occurring in model training, before the model training, the continuity characteristic of the data set is subjected to standardization processing, so that the continuity characteristic of the data set is in standard normal distribution, and the possibility that the model falls into a linear region of a nonlinear activation function due to an outlier is reduced. The normalization of the continuum profile is shown in equation (3), where x' represents the input data, μ represents the mean of the input data, and S represents the standard deviation of the input data.
In addition, in order to make the distribution of the model label data more stable, reduce the calculated amount of the model, and avoid introducing new calculation errors due to the difference between the data distribution in the training set and the data distribution in the test set, we choose to perform logarithmic processing on the charging load data y' of the station, as shown in formula (4).
y＝log(y′+1) (4)
Specifically, a comparison of the charging load data of the logarithmic process is shown in fig. 2 and fig. 3, where fig. 2 is a station charging load curve before the logarithmic process, and fig. 3 is a station charging load curve after the logarithmic process.
Experimental data a total of 959040 sample examples were randomly selected in this example to cover 22 days, where the test set samples were the data of the last day in the data set, and the training set and the validation set were the data of the first 21 days in the data set, as follows: 2, the verification set is separated from the training data to facilitate selection of a better model in multiple training, and the final separation result is shown in table 3.
TABLE 3 Experimental data set partitioning results
3. Model training
The overall structure diagram of the model is shown in fig. 4, and the present embodiment will explain the structure of the model by composing the input and output of the model and the modules of the model into two parts.
3-1 model input and output
The input of the prediction model is divided into two parts, and if the current predicted content is the load value of the time period t, the first part of the model input data is x shown on the left side of fig. 4 t ＝[x t0 ，x t1 ，...，x td ]，x t ∈R 1×d And h t ＝[h t0 ，h t1 ，...，h td ]，h t ∈R 1×k Respectively representing a continuous characteristic vector and a category characteristic vector corresponding to the time period t; the second part being a right-hand historical data encoding module
u t ∈R s×c And
p t ∈R s×1 the historical data step length s =7 is obtained in our experiment, that is, the characteristic data and the charging load value corresponding to the time period t in the past 7 days are taken as the input of the historical data coding module.
Correspondingly, the output of the prediction model is the load value corresponding to the t time period of the day predicted by the model
3-2 model each module composition
The model mainly comprises three modules: 1) A feature embedding module; 2) A historical data encoding module; 3) And a fusion prediction module.
The continuous type characteristic and the classification type characteristic are separately input in the characteristic embedding module, and the classification type characteristic h t Input Embedding layer converts category features into continuous type feature vector h' t ∈R 1×m Where m = k _ embedding _ size, in the present embodiment, embedding _ size =3 is taken, and the output result h 'after the embedding process is used as the type feature' t ∈R 1×3k And continuous type feature x t Combine to obtain a new vector r t ＝[h′ t |x t ]，r t ∈R 1×e ，e＝d+3k。
Input of historical data encoding Module model input data u t 、p t And model other Module output data g t ∈R 1 ×(j-t) Composition of g t Is an amount r t The part of the feature vector related to the medium time and the weather is obtained by means of vector slicing, and i and j can be defined as an expression (5) according to the feature arrangement condition.
g t ＝r t [：，i：j] (5)
As can be seen in FIG. 4, the historical data encoding module first encodes u t Input into a network infrastructure module FA, the structural diagram inside the network infrastructure module FA is shown in the upper right corner of fig. 4, where Linear represents a Linear layer, which is defined as shown in equation (6), where W and b represent a weight matrix and an offset that can be learned, respectively, where X = [ X ], [ 0 ，X 1 ，...，X i ，..]In order to realize the jump addition method (i.e. the residual connection method) between layers in the historical data encoding module of the method of the embodiment, we set the output dimension and the input dimension of the linear layer of the module to be equalThe value is obtained.
f(X)＝XW T +b (6)
The BatchNorm layer in the FA block is defined as shown in equation (7), where μ represents the expectation of the input data X, σ is the standard deviation, and e is a value close to 0, and γ and β represent learnable affine transformation parameter vectors with the same vector dimension as that of the input data in the layer in order to avoid calculation errors when the standard deviation σ = 0.
Tanh is the activation function, as shown in equation (8):
to u in historical data coding module t Is performed in a residual concatenation manner, which can be expressed as formula (9), where
K t ∈R s×z Z is a hyper-parameter of the model, which can be adjusted manually according to a priori knowledge, in this embodiment z =16.
K＝FA(u t +FA(u t )) (9)
For another input g of the module t Is obtained after being processed by the same residual error connection mode
Q t ∈R z×1 。
Q t ＝FA(g t +FA(g t )) (10)
Then K is put t And Q t After multiplication, normalization processing is carried out to obtain t time period in historical dataThe value of the influence factor alpha on the t time period predicted on the day t ∈R s×1 It is calculated as shown in equation (11), where the definition of the softmax function is given by equation (12).
α t ＝softmax(K t ×Q t ) (11)
Finally using alpha t To p t Weighting and calculating to obtain output vector V of historical data encoding module t ∈R 1×1 As shown in equation (13):
input data O of fusion prediction module t Obtained by combining the output of the feature embedding module and the output of the historical data encoding module, i.e. O t ＝[r t ，V t ]，O t ∈R 1×(e+1) 。
The fusion prediction module also adopts a residual error connection mode, and is different from the historical data coding module in that the output dimension and the input dimension of a linear layer in the network basic module FA are different, so that a characteristic combination mode rather than a summation mode is adopted in the residual error connection, the last part of the module, namely a Predict block in a figure 4, consists of 2 FA blocks and one linear layer, namely the last layer of the whole model is the linear layer, the output dimension of the linear layer is 1, and the output content is a logarithmic load value obtained by model calculation
The modular calculation process can be expressed as equation (14), where W p ∈R c×1 ，b p ∈R 1×1 And c denotes the dimension of the last FA block output vector.
Finally, regarding model loss function selection, the task in this embodiment belongs to a regression task, and the objective is to fit a load curve, so we adopt a Mean-Squared Error (MSE) loss function commonly used in the regression task, and the formula definition is shown in formula (15). In the method, error back propagation is realized and the weight of the network is updated through a random Gradient Descent (SGD) + Momentum (Momentum) minimum loss function, the initial learning rate of model training is set to be 0.12, and the learning rate adjustment strategy is ReduceLROnPlateau.
Wherein, y t The actual charging load value corresponding to the time t obtained from the history data.
2. Model prediction phase
Similarly, the model prediction phase, again consists of a number of steps: 1. preparing data; 2. model prediction; 3. and (6) restoring data.
1. Data preparation
The data acquisition and storage are configured in the model prediction stage, and only the normal operation of the data acquisition and storage program is required to be ensured, so that the operation is not required to be repeated in the prediction stage, the processing mode of the characteristics is the same as that in the model training stage, and the weather-related characteristics are future weather prediction results obtained through a weather forecast interface.
2. Model prediction
In the model prediction process, two cases are divided: 1) Predicting the load value of the next day of the site; 2) And predicting the load value of the station on the day.
In the first case, we basically adopt a way to predict the charging load value of a station corresponding to 1440 minutes per minute for one station at a time, in which case, the current day is not over, so the historical data of the historical data coding module of the model is shifted forward by one day. In the second case, the prediction time range may be selected, and the load value of the actually generated time period does not need to be predicted by the model, that is, the prediction result is corrected by the real data before the current time period, so that the inference time of the model may be effectively reduced, if we need to predict the charging load curve of the station on the same day, each station should correspond to 1440 prediction samples, if the current time period is at the t-th minute of the day, the [0, t-1] time period belongs to the generated time period, the charging load value of the station in the corresponding time period should be replaced by the real value, and the model actually predicts only the load value in the [ t,1439] time period when inferring, so that only the data input to the model is also the feature vector of the [ t,1439] time period.
3. Data recovery
In order to make the data distribution of model labels more stable, reduce the calculated amount of the model and avoid introducing new calculation errors due to the data distribution difference of a training set in a test set, the charging load data of a station is subjected to logarithmic processing before the data is input into the model, the real load value can not be restored in the training stage, but the real load value needs to be restored in the prediction stage, and the data restoration mode is shown in formula (16), wherein y' represents the output value of the model, and y represents the restored charging load.
y＝e y′ -1 (16)
Based on the charging load prediction method, in order to verify the prediction accuracy and the improvement compared with the prior art, the evaluation index is selected as R in the embodiment 2 (R-Squared), which is also called fitness and is commonly used to measure the fit of the predicted result to the real situation, is generally defined as shown in equations (17) - (20), where SST (Total Sum of Squares) is the sample variance, SSE (Sum of Squares of Error) is the residual Sum of Squares, and SSR (Sum of Squares of the Regression) is the Regression Sum of Squares.
As shown in table 4, the experimental results show that the prediction model provided in this embodiment can achieve a better effect on the prediction of the charging load of the charging station of the electric vehicle, and the average R of each station in the test set 2 0.72, up to 0.95, can be used to provide data support for the construction planning and daily operation of the charging station. The reference models obtained by removing and replacing the modules on the basis of the model provided by the invention are numbered 3-6 in the table 4, wherein MLP is a model for removing a residual error connection mode and a historical data coding module, MLP _ RES is a model for removing a historical data coding module, MLP _ ATT is a model for removing a residual error connection mode, and MLP _ RES _ GRU is a model for replacing an attention coding layer in the historical coding module by a GRU. From the experimental results in table 4, it can be seen that the LightGBM, XGBoost, MLP _ RES have no historical data processing module, and the model fitting effect is poor; the model effect of removing and replacing part of the modules is slightly worse than that of the method proposed by the inventor, which also shows that the modules of the proposed model are effective.
TABLE 4R of control and experimental models on test set 2
By adopting the charging load prediction method, the four randomly selected stations are subjected to prediction verification, and the real charging loads of the four stations and the prediction results are compared with a graph shown in fig. 5 to 8.
Claims (10)
1. A charging station charging load prediction method based on an attention mechanism is characterized by comprising a prediction model, wherein the prediction model comprises a feature embedding module, a historical data coding module and a fusion prediction module;
the prediction model training steps are as follows:
acquiring characteristic data of a charging station time period t and a charging load value y corresponding to the time period t t All feature data constituting a feature vector r t (ii) a Extracting a feature vector r t The part related to the middle time and the weather characteristics is recorded as g t ；
Acquiring historical data of a charging station in a period t of the past s days, including
And
respectively representing the date, time and weather characteristics of a time period t in the past s days and the load value in the corresponding time period;
the input vector of the historical data encoding module is u t And g t ；
The historical data encoding module is provided with a network basic module FA which is sequentially provided with a linear layer, a BatchNorm layer, a Tanh layer and a Dropout layer from input to output;
inputting the vector X into a network basic module FA to obtain FA (X);
historical data encoding module pair vector u t The process of (3) is a residual concatenation process:
K t ＝FA(u t +FA(u t ))，
for g to t The same treatment gives: q t ＝FA(g t +FA(g t ))；
K t And Q t After multiplication, normalization processing is carried out to obtain the prediction y of the characteristic data pair corresponding to the time period t in the historical data t Influence factor alpha of t :
α t ＝softmax(K t ×Q t )
By alpha t To p t Weighting and calculating to obtain the output vector V of the historical data coding module t ：
The input vector of the fusion prediction module is O t ，O t ＝[r t ,V t ]；
The output of the fusion prediction module is:
wherein, W p And b p Respectively representing a weight matrix and an offset, wherein the weight matrix and the offset are updated weights after learning and training in a prediction model;
setting the mean square error loss function L MSE ，
Performing error back propagation and updating the prediction model based on a random gradient descent algorithm and a momentum minimization loss algorithm until the prediction model converges to obtain a converged prediction model;
the prediction steps of the prediction model are as follows:
recording the time period to be predicted as t ', acquiring characteristic data of the time period t', wherein the characteristic data related to weather is acquired through weather forecast, and substituting the characteristic data into the converged prediction model for calculation to obtain the characteristic data
I.e. the predicted charging load value for time period t'.
2. The charging station charging based on attention mechanism as claimed in claim 1The load prediction method is characterized in that the characteristic data comprises x t ＝[x t0 ,x t1 ,…,x td-1 ]And h t ＝[h t0 ,h t1 ,…,h tk-1 ]，x t A continuous feature vector h formed by d continuous feature data corresponding to the representation time interval t t A category type feature vector composed of k category type feature data corresponding to the time period t; class type feature vector h t Converting input feature embedding module into continuous feature vector h' t Continuous type feature vector x t And continuous type feature vector h' t Combining to obtain the characteristic vector r t ＝[h′ t |x t ]。
3. The method of claim 2, wherein the feature embedding module processes the input class-type feature vector as follows: one-hot coding is carried out on each feature, and then a learnable weight matrix is multiplied to convert the feature into a continuous feature vector with dimension r, wherein r is a hyper-parameter of the model and is determined through priori knowledge and manual debugging.
4. The method of claim 1, wherein the linear layer has an input-output relationship of:
f(X)＝XW T +b，
x represents an input vector of the layer, W and b represent a weight matrix and an offset respectively, and the weight matrix and the offset are updated weights after learning and training are needed in a prediction model;
the input and output relational expression of the BatchNorm layer is as follows:
mu represents the expectation of an input vector X, sigma is a standard deviation, epsilon is a positive number close to 0, gamma and beta respectively represent affine transformation parameter vectors and are weight values which need to be updated after learning and training in a prediction model;
the input and output relational expression of the Tanh layer is as follows:
5. the method as claimed in claim 1, wherein the feature data input into the prediction model is normalized by the following formula:
x' represents raw data, μ represents a mean value of the raw data, S represents a standard deviation of the raw data, and x is data after normalization processing.
6. The method of claim 1, wherein the charging load value is a logarithmized charging load value, and the logarithmized charging load value is represented by the following formula:
y＝log(y′+1)
where y' represents the original data and y represents the data after the logarithmic process.
7. The attention-based charging station charging load prediction method of claim 1, wherein the softmax function is formulated as:
8. the method as claimed in claim 1, wherein the characteristic data at least includes charging station information, location information of charging station, weather information, and time information.
9. The method of claim 8, wherein the weather information comprises a current temperature and a current humidity.
10. The method of claim 8, wherein the time information comprises hours and minutes of a day that the current time is.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211188648.7A CN115526398A (en) | 2022-09-27 | 2022-09-27 | Attention mechanism-based charging load prediction method for charging station |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211188648.7A CN115526398A (en) | 2022-09-27 | 2022-09-27 | Attention mechanism-based charging load prediction method for charging station |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115526398A true CN115526398A (en) | 2022-12-27 |
Family
ID=84699633
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211188648.7A Pending CN115526398A (en) | 2022-09-27 | 2022-09-27 | Attention mechanism-based charging load prediction method for charging station |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115526398A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117310277A (en) * | 2023-09-25 | 2023-12-29 | 国网四川省电力公司营销服务中心 | Electric energy metering method, system, equipment and medium for off-board charger of electric automobile |
-
2022
- 2022-09-27 CN CN202211188648.7A patent/CN115526398A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117310277A (en) * | 2023-09-25 | 2023-12-29 | 国网四川省电力公司营销服务中心 | Electric energy metering method, system, equipment and medium for off-board charger of electric automobile |
| CN117310277B (en) * | 2023-09-25 | 2024-06-04 | 国网四川省电力公司营销服务中心 | Electric energy metering method, system, equipment and medium for off-board charger of electric automobile |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108539738B (en) | Short-term load prediction method based on gradient lifting decision tree | |
| CN107451686B (en) | Micro-grid energy optimization method of genetic algorithm considering random prediction error | |
| CN109002942A (en) | A kind of short-term load forecasting method based on stochastic neural net | |
| Carriere et al. | A novel approach for seamless probabilistic photovoltaic power forecasting covering multiple time frames | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN110212524A (en) | A kind of region Methods of electric load forecasting | |
| CN111178585A (en) | Fault reporting amount prediction method based on multi-algorithm model fusion | |
| CN112990587B (en) | Method, system, equipment and medium for accurately predicting power consumption of transformer area | |
| CN110619442A (en) | Vehicle berth prediction method based on reinforcement learning | |
| CN116090602A (en) | Power load prediction method and system | |
| CN115965132A (en) | Distributed photovoltaic digital twin system power prediction method based on GA-BP neural network | |
| CN106919645A (en) | The sight spot meteorological element Intelligent fine Forecasting Methodology at the big scenic spot of complex landform | |
| CN115526398A (en) | Attention mechanism-based charging load prediction method for charging station | |
| CN116599151A (en) | Source network storage safety management method based on multi-source data | |
| CN111915083A (en) | Wind power prediction method and prediction system based on time hierarchical combination | |
| CN114665467A (en) | Intelligent scheduling method and system for light storage and charging micro-grid system | |
| CN115760213A (en) | Method for constructing short-term electricity price prediction model, short-term electricity price prediction method, computer-readable medium, and electronic device | |
| CN116646929A (en) | PSO-CNN-BILSTM-based short-term wind power prediction method | |
| CN115062856A (en) | New energy plant station medium-term output prediction method and system based on power grid private cloud | |
| CN114444805A (en) | Control method for smooth output of multi-photovoltaic power station shared energy storage system | |
| CN117787915A (en) | Digital twin intelligent brain construction method for power distribution network | |
| CN116227738B (en) | Method and system for predicting traffic interval of power grid customer service | |
| CN117473461A (en) | Urban charging load space-time prediction method based on voltage class classification | |
| CN117458481A (en) | XGBoost-based power system load prediction method and device | |
| CN117410959A (en) | EEMD-GWO-LSTM network and MC error correction-based power grid load ultra-short-term prediction method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |