CN103400203A - Electric vehicle charging station load prediction method based on support vector machine - Google Patents

Abstract

The invention provides an electric vehicle charging station load prediction method based on a support vector machine. The electric vehicle charging station load prediction method based on the support vector machine comprises the following steps of: collecting charging station load before a prediction day and historical data of factors which affect the charging station load; constructing a training sample set according to the historical data; training by the training sample set to obtain the Lagrange multiplier of a support vector machine regression function; according to the Lagrange multiplier, establishing a support vector machine prediction model; and substituting the prediction sample set into the support vector machine prediction model to obtain a charging station load prediction value. According to the electric vehicle charging station load prediction method based on the support vector machine, which is disclosed by the invention, the electric vehicle charging station load prediction precision can be effectively improved, and a more reliable basis is provided for an electric power dispatching department to regulate a dispatching plan, regulate system reserve capacity and optimize the generator set capacity.

Description

A kind of load forecasting method of electric automobile charging station based on support vector machine

Technical field

The invention belongs to electric automobile charging station load prediction field.

Background technology

The tradition fuel-engined vehicle not only consumes non-renewable energy resources in a large number, is also one of main source of Present Global greenhouse gas emission and the local pollution of the environment.Because electric automobile replaces fuel oil with electric power, playing great function aspect the solution energy and environmental crisis, electric automobile just progressively moves towards the large-scale application stage by development.But that large-scale electric vehicle charging electric load has is high-power, undulatory property and the characteristics such as uncertain, and this controls for safe operation, Optimized Operation and the trend of electrical network and has all brought very large challenge.Therefore predict significant to the load of electric automobile charging station.

After electric automobile is popularized on a large scale, its charging load will be very large, power scheduling department can be according to predicting the outcome that electric automobile charging station is loaded, adjust operation plan, adjustment System margin capacity, optimize genset and exert oneself etc., can reduce operating cost, strengthen security, reliability and the controllability of system.

At present, the electric automobile charging station load prediction is according to electric automobile during traveling characteristic and battery charging/discharging characteristic.At first predict that according to the characteristics of travelling of electric automobile it arrives time and the battery dump energy of charging station charging, utilize the charging characteristic curve of these data and battery to determine each charge power constantly of every electric automobile, and then obtain total charge power of charging station.This fado is used for the electric automobile charging station initial operation stage, and historical load information is less, and the data of grasp only have the situation of charge-discharge characteristic of time, mileage and the battery of electric automobile operation.Present charging station load forecasting method substantially all is based on this method.Yet Vehicle Driving Cycle has very large randomness and uncontrollability, causes vehicle, to time and the battery dump energy of charging station charging, very large randomness is arranged, and therefore, this method precision that predicts the outcome is lower.

Summary of the invention

The object of the present invention is to provide a kind of load forecasting method of electric automobile charging station based on support vector machine, be intended to solve tradition and only utilize the not enough problem of precision of the charging station load forecasting method of electric automobile during traveling characteristic and battery charge characteristic prediction load, improve adaptivity and the engineering practicability of electric automobile charging station load forecasting method.

For achieving the above object, the present invention has adopted following technical scheme:

A kind of load forecasting method of electric automobile charging station based on support vector machine, this Forecasting Methodology comprises the following steps:

1) charging station load and the historical data that affect factor of charging station load before Collection and Forecast Day, described factor comprise all attributes, red-letter day attribute, temperature, weather condition, charging station service vehicle number and the charging set number that works;

2) each affect the value of the factor of charging station load to determine prediction day, and wherein, temperature and weather condition determine according to weather forecast, and the several data that provide according to charging station of charging station service vehicle number and normal operation charging set are definite;

3) to historical data and step 2) value determined makes normalized;

4) after step 3), historical data and step 2 that utilization and prediction day k are constantly corresponding) definite value structure training sample set and forecast sample collection;

5) utilize the training sample set training to obtain the Lagrange multiplier of the support vector machine regression function constantly corresponding with k, according to described Lagrange multiplier, set up the SVM prediction model constantly corresponding with k;

6) forecast sample collection substitution SVM prediction model is obtained the charging station load prediction value constantly corresponding with k;

7) repeating step 4), to step 6), obtain predicting a day charging station load prediction value corresponding to each moment.

Described all attributes are designated as W, W ∈ { 1,2,3,4,5,6,7}, W=1 corresponding Monday, W=2 corresponding Tu., W=3 corresponding Wednesday, W=4 corresponding Thursday, W=5 corresponding Friday, W=6 corresponding Saturday, W=7 corresponding Sun.; Red-letter day, attribute was designated as F, and { 0,1}, if get F=1 red-letter day, otherwise get F=0 to F ∈; Weather condition is designated as A, A ∈ 1,2,3}, and the A=1 correspondence is fine, and A=2 is corresponding cloudy, the corresponding sleet of A=3.

Described historical data for before prediction day in 100-200 days the electric automobile charging station data-base recording about the charging station load and the data message that affect factor that charging station loads.Determine the number of times of prediction in a day according to the step-length of prediction, all use the data message of corresponding moment correspondence to train at every turn, obtain the SVM prediction model in this moment, and with the charging station load value in this this moment of model prediction (when for example prediction step is 10 minutes, can obtain 144 SVM prediction models every day, the charging station load prediction value in measurable 144 moment).

Before historical data is made normalized, the historical data that gathers is scanned, if when in historical data, one day, charging station load constantly was for negative value, with modified value, replace negative value, modified value is the mean value of forward and backward two days of described one day corresponding charging station constantly load.

The construction method of described training sample set is: the front 10 days k charging station load value L={L constantly that gets i day _{I-1, k}, L _{I-2, k}, L _{I-3, k}, L _{I-4, k}, L _{I-5, k}, L _{I-6, k}, L _{I-7, k}, L _{I-8, k}, L _{I-9, k}, L _{I-10, k}, all attribute W of i day _i, i day attribute F in red-letter day _i, i day temperature T _i, i day weather condition A _i, i day charging station service vehicle counts B _iAnd the charging set that works i day is counted C _iInput vector as i day k training sample constantly:

x _i,k={L _i-1,k,L _i-2,k,L _i-3,k,L _i-4,k,L _i-5,k,L _i-6,k,L _i-7,k,L _i-8,k,L _i-9,k,L _i-10,k,W _i,F _i,T _i,A _i,B _i,C _i}

Get i day k charging station load L constantly _i,kTarget output value as training sample:

y _i,k=L _i,k

Can obtain by that analogy input vector and the target output value of the training sample of day any time, the training sample set of any time arbitrarily

I=11,12 ..., N, N are the collection number of days of historical data.

The construction method of described forecast sample collection is: the front 10 days k charging station load value L={L constantly that gets prediction day (M day) _{M-1, k}, L _{M-2, k}, L _{M-3, k}, L _{M-4, k}, L _{M-5, k}, L _{M-6, k}, L _{M-7, k}, L _{M-8, k}, L _{M-9, k}, L _{M-10, k}, the prediction day all attribute W _M, the prediction day attribute F in red-letter day _M, the prediction day temperature T _M, the prediction day weather condition A _M, a prediction day charging station service vehicle counts B _MAnd a prediction day normal operation charging set is counted C _MAs the input vector of k forecast sample constantly, M=N+1:

x _k={L _M-1,k,L _M-2,k,L _M-3,k,L _M-4,k,L _M-5,k,L _M-6,k,L _M-7,k,L _M-8,k,L _M-9,k,L _M-10,k,W _M,F _M,T _M,A _M,B _M,C _M}

Obtain by that analogy the forecast sample collection of any time.

The mathematical model that be used for Support Vector Machines Optimized parameter constantly corresponding with k is:

$\max \underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}-\frac{1}{2}\underset{i=11}{\overset{N}{\mathrm{\Σ}}}\underset{j=11}{\overset{N}{\mathrm{\Σ}}}{y}_{i,k}{y}_{j,k}{\mathrm{\α}}_{i,k}{\mathrm{\α}}_{j,k}(\mathrm{\φ}\left(x_{i,k}\right)\·\mathrm{\φ}\left(x_{j,k}\right))$

$s.t.\underset{i-11}{\overset{N}{\mathrm{\Σ}}}{y}_{i,k}{\mathrm{\α}}_{i,k}=0$

0≤α _i,k≤C,i=11,12,...,N

α _i,kAnd α _j,kThe mathematical model i that is used for the Support Vector Machines Optimized parameter and j Lagrange multiplier that expression is constantly corresponding with k, C represents penalty factor.

The SVM prediction model constantly corresponding with k is:

$f\left(x_k\right)=\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_k\right)\·\mathrm{\φ}\left(x_{i,k}\right))+y_{j,k}-\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_{i,k}\right)\·\mathrm{\φ}\left(x_{j,k}\right))$

Wherein, x _kThe input vector of expression forecast sample, N represents the collection number of days of historical data,

The optimal value of i the Lagrange multiplier that the expression basis mathematical model optimizing that be used for Support Vector Machines Optimized parameter corresponding with the k moment obtains, x _j,kConcentrate for k training sample constantly, be positioned at the input vector of support vector position j, y _j,kFor k training sample is constantly concentrated, be positioned at the target output value of support vector position j, φ () is illustrated under input and the inseparable condition of output linearity, and lower dimensional space is to the mapping of higher dimensional space, φ (x) can not occur separately, but with inner product φ (x _i,k) φ (x _j,k) form occur,, therefore model does not need to know the concrete form of φ (x), only need know its inner product φ (x _i,k) φ (x _j,k) form get final product, note φ (x _i,k) φ (x _j,k)=K (x _i,k, x _j,k), φ (x _k) φ (x _i,k)=K (x _k, x _i,k), K (x _i,k, x _j,k) and K (x _k, x _i,k) express support for the kernel function that vector machine adopts.

Due to the high dimension (16 dimension) of input sample, described kernel function adopts gaussian kernel function, and for example concrete form is K (x _i,k, x _j,k)=exp (|| x _i,k-x _j,k|| ²/ 2 σ ²).

Described Forecasting Methodology is further comprising the steps of: call figure interface display instrument, the form of prediction day each charging station load prediction value constantly with curve represented.

Beneficial effect of the present invention is embodied in:

the present invention utilizes support vector machine (SVM) method, factor (all attributes of traditional impact load have been considered, red-letter day attribute, temperature, weather condition) and the new influence factor of bringing after electric automobile access (the electric automobile number of charging station service, the charging set number of charging station normal operation), each charging station load constantly of prediction electric automobile charging station, can effectively improve the precision of electric automobile charging station load prediction, for power scheduling department adjusts operation plan, the adjustment System margin capacity, optimizing genset exerts oneself etc. more reliable foundation is provided.

In order further to reduce the impact of non-True Data on predicated error in historical data, the present invention first carries out pre-service to historical data, replaces the negative value in historical charging station load, and the data that recycling was processed are trained and predict.Result shows, is taking the pre-service front and back, and the mean absolute error of prediction is reduced to 103.836 by 166.402, and root-mean-square error is reduced to 0.101 by 0.141.

The present invention adopts heuritic approach, and when historical data is 100-200 days, training sample set and forecast sample while concentrating historical load to choose the charging station load of front 10 days, the charging station load error that prediction obtains is less, and computing velocity is very fast.

Description of drawings

Fig. 1 is electric automobile charging station load forecasting method process flow diagram of the present invention;

Fig. 2 is the electric automobile charging station load prediction curve.

Embodiment

The invention will be further described below in conjunction with accompanying drawing.

The present invention utilizes support vector machine (SVM) method, consider the charging load of the information prediction electric automobile charging stations such as charging set number, temperature, weather information and association attributes of electric automobile number, the charging station normal operation of the service of electric automobile charging station historical load, charging station, method of this prediction electric automobile charging station charging load is specific as follows:

1) sample chooses

Identical with conventional electric power load, the electric automobile charging station load also has certain periodicity and continuity, and different is, and charging load value and the electric automobile of charging station service and the quantity of charging set have obvious relation., by the analysis to the charging station historical data, determine that the input vector of training sample set and forecast sample collection comprises: certain is the charging station load value in the front 10 day corresponding moment constantly; All attributes on the same day, red-letter day attribute; The temperature on the same day, weather condition; The same day charging station service vehicle number and normal operation charging set number.

The present invention is used the prediction historical data of 150 days a few days ago as training sample, and training obtains after the SVM model, the electric automobile charging station load of prediction day being predicted.

2) correction of historical sample (pre-service)

Failure and other reasons due to electric automobile charging station data acquisition communication system, make the power (i.e. load) of the part charging set that collects be negative value, and then cause having error between the historical data of charging station historical data library storage and actual, historical data value, namely there be " bad point " (being non-True Data) in historical data.In order to prevent the precision of these " bad point " impact prediction results, the present invention revises " bad point ".Because charging station load has certain periodicity, so the present invention is with the modified value of the mean value conduct " bad point " of synchronization charging station load on the two before and after it.

3) to the sample normalized

The order of magnitude and the variation range of each sample data differ greatly, and in order to prevent the less data of variation range, by variation range larger data submerge and computation process, are overflowed, and the present invention carries out normalized to training sample set and forecast sample collection.

4) sample training and prediction

Utilize training sample set to train the Lagrange multiplier that obtains with each constantly corresponding support vector machine regression function, according to the SVM prediction model that Lagrange multiplier is set up and each is constantly corresponding; Forecast sample collection substitution SVM prediction model is obtained and each constantly corresponding charging station load prediction value;

Setting up rational SVM model is gordian technique of the present invention, and the process of specifically setting up of this model is as follows:

4a) the ultimate principle of support vector machine (SVM)

A given training sample set

x _iFor input vector, y _iFor target output value, n is the number of training sample, and support vector machine returns and can be used to find input vector x _iWith corresponding output valve y _iBetween relation.When the pass between sample output and input vector while being non-linear, SVM is not the relation of directly finding input and output, but first by Nonlinear Mapping φ (x): R ^N→ R ^mThe input space is mapped to high-dimensional feature space, and then realizes linear regression in high-dimensional feature space:

f(x)=(w,φ(x))+b (1)

In formula (1), w is weighted value, and b is deviation.SVM adopts structural risk minimization principle, and w and b are trained by formula (2):

$\min \frac{{\left|\right|w\left|\right|}^2}{2}+c\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i$

s.t.y _i[(w·φ(x) _i)+b]≥1-ξ _i (2)

ξ _i≥0 i=1,...n

In the objective function of formula (2), first is the regularization part, and second is the experience error.Coefficient C is penalty factor.The Lagrangian function of structure following formula, result is:

$L(w,b,\mathrm{\ξ})=\frac{{\left|\right|w\left|\right|}^2}{2}+c\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i(y_i(w\·\mathrm{\φ}{\left(x\right)}_i)+b-1+{\mathrm{\ξ}}_i)-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i{\mathrm{\ξ}}_i---\left(3\right)$

α=(α wherein ₁..., α _n), β=(β ₁..., β _n), α _i〉=0, β _i〉=0, i=1,2 ... n.

According to KKT(Karush-Kuhn-Tucker) condition, ask respectively formula (3) to w, b and ξ _iPartial differential, will obtain w, b, α _i, β _iAnd ξ _iRelation:

$w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i\mathrm{\φ}\left(x_i\right)---\left(4\right)$

${b=y}_j-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i(\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right))$

In formula (4) substitution formula (3), can be with the problem reduction of formula (2), and application Wolfe antithesis skill is (referring to A.Smola and B.Scholkopf, A tutorial on support vector regression, NeuroCOLT Tech.Rep.TR1998-030, Royal Holloway College, London, U.K., 1998.) be converted into its dual problem:

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{y}_i{y}_j{\mathrm{\α}}_i{\mathrm{\α}}_j(\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right))$

$s.t.\underset{i-1}{\overset{n}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i=0---\left(5\right)$

0≤α _i≤C i=1,2,...,n

In formula (5), introduce kernel function K (x _i, x _j)=(φ (x _i) φ (x _j)), thereby only need determine function K (x _i, x _j),, in the situation that need not determine φ (x) expression, can obtain corresponding decision function.Application formula (5) is trained training sample set, obtains under optimal condition

Utilize formula (4) to try to achieve w ^*, b ^*,, with its substitution formula (1), get final product to obtain the SVM regression function:

$f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{*}{y}_i(\mathrm{\φ}\left(x\right)\·\mathrm{\φ}\left(x_i\right))+y_j-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{*}{y}_i(\mathrm{\φ}\left(x_i\right)\·\mathrm{\φ}\left(x_j\right))---\left(6\right)$

4b) determine the input sample of support vector machine (SVM)

Identical with the conventional electric power load, the electric automobile charging station load also has certain periodicity and continuity, and different is that charging load value and the electric automobile quantity of operation and the quantity of charging set etc. have obvious relation.The present invention, by the analysis to the charging station historical data, determines that the input vector of training sample set and forecast sample collection comprises:

1) be the periodicity of performance charging load, choose certain front 10 days constantly corresponding charging load value L={L constantly ₁, L ₂, L ₃, L ₄, L ₅, L ₆, L ₇, L ₈, L ₉, L ₁₀;

2) all attribute W ∈ { 1,2,3,4,5,6,7} on the same day;

3) the attribute F ∈ in red-letter day on the same day 0,1}(gets F=1 red-letter day, otherwise gets F=0);

4) the weather condition A ∈ on the same day the fine F=1 that gets of 1,2,3}(weather, the moon is got F=2, sleet is got F=3);

5) temperature T on the same day;

6) the vehicle number B of planning operation on the same day;

The charging set that 7) can normally use the same day is counted C.

Comprehensive above seven parts can obtain the training sample set of support vector machine (SVM) and the input vector of forecast sample collection: 16 dimension input vector { L ₁, L ₂, L ₃, L ₄, L ₅, L ₆, L ₇, L ₈, L ₉, L ₁₀, W, F, A, T, B, C}.

4c) the SVM model is established and performing step

The quality of SVM forecast model depends on selected kernel function and parameter to a great extent.Input dimension of the present invention higher (16 dimension), number of training is very large simultaneously, and at the premium properties that solves on the higher-dimension large sample problem, the present invention chooses gaussian kernel function and sets up the SVM model based on gaussian kernel function, and parameter is chosen as C=100.

After having set up the SVM model, use the prediction above-mentioned historical information of 150 days a few days ago as training sample, training obtains after the SVM forecast model, the electric automobile charging station load being predicted.The specific implementation process as shown in Figure 1.

1) charging station load and the historical data that affect factor of charging station load before Collection and Forecast Day, described factor comprise all attributes, red-letter day attribute, temperature, weather condition, charging station service vehicle number and the charging set number that works;

The historical data of 2) charging station that gathers being loaded scans, if when in historical data, one day, charging station load constantly was for negative value, with modified value, replace negative value, modified value is the mean value of charging station load of the corresponding moment of forward and backward two days of described one day;

3) determine that prediction day, each affected the value of the factor of charging station load, wherein, temperature and weather condition (are for example determined according to weather forecast, temperature is got the maximal value of forecast), the several data that provide according to charging station of charging station service vehicle number and normal operation charging set (are for example determined, charging station service vehicle number is definite by the intended services vehicle number that charging station receives, and normal operation charging set number equals the charging set number that works the previous day);

4) historical data and the definite value of step 3) are made normalized;

5) after step 4), the value of utilizing the historical data constantly corresponding with prediction day k and step 3) to determine builds training sample set and forecast sample collection;

6) utilize the training sample set training to obtain the Lagrange multiplier of the support vector machine regression function constantly corresponding with k, according to described Lagrange multiplier, set up the SVM prediction model constantly corresponding with k;

7) forecast sample collection substitution SVM prediction model is obtained the charging station load prediction value constantly corresponding with k;

8) repeating step 5), to step 7), obtain predicting a day charging station load prediction value corresponding to each moment.

9) after prediction is completed, call figure interface display instrument, the form of prediction day each charging station load prediction value constantly with curve represented, simultaneously each predicted value constantly is stored in database.

The construction method of training sample set is: the front 10 days k charging station load value L={L constantly that gets i day _{I-1, k}, L _{I-2, k}, L _{I-3, k}, L _{I-4, k}, L _{I-5, k}, L _{I-6, k}, L _{I-7, k}, L _{I-8, k}, L _{I-9, k}, L _{I-10, k}, all attribute W of i day _i, i day attribute F in red-letter day _i, i day temperature T _i, i day weather condition A _i, i day charging station service vehicle counts B _iAnd the charging set that works i day is counted C _iInput vector as i day k training sample constantly:

x _i,k={L _i-1,k,L _i-2,k,L _i-3,k,L _i-4,k,L _i-5,k,L _i-6,k,L _i-7,k,L _i-8,k,L _i-9,k,L _i-10,k,W _i,F _i,T _i,A _i,B _i,C _i}

Get i day k charging station load L constantly _i, _kTarget output value as training sample:

y _i,k=L _i,k

Can obtain by that analogy input vector and the target output value of the training sample of day any time, the training sample set of any time arbitrarily

K=1,2...144, i=11,12 ..., N, N are the collection number of days of historical data.

The construction method of forecast sample collection is: the front 10 days k charging station load value L={L constantly that gets prediction day (M day) _{M-1, k}, L _{M-2, k}, L _{M-3, k}, L _{M-4, k}, L _{M-5, k}, L _{M-6, k}, L _{M-7, k}, L _{M-8, k}, L _{M-9, k}, L _{M-10, k}, the prediction day all attribute W _M, the prediction day attribute F in red-letter day _M, the prediction day temperature T _M, the prediction day weather condition A _M, a prediction day charging station service vehicle counts B _MAnd a prediction day normal operation charging set is counted C _MAs the input vector of k forecast sample constantly, M=N+1:

x _k={L _M-1,k,L _M-2,k,L _M-3,k,L _M-4,k,L _M-5,k,L _M-6,k,L _M-7,k,L _M-8,k,L _M-9,k,L _M-10,k,W _M,F _M,T _M,A _M,B _M,C _M}

Obtain by that analogy the forecast sample collection of any time.

The mathematical model that be used for Support Vector Machines Optimized parameter constantly corresponding with k is:

$\max \underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}-\frac{1}{2}\underset{i=11}{\overset{N}{\mathrm{\Σ}}}\underset{j=11}{\overset{N}{\mathrm{\Σ}}}{y}_{i,k}{y}_{j,k}{\mathrm{\α}}_{i,k}{\mathrm{\α}}_{j,k}(\mathrm{\φ}\left(x_{i,k}\right)\·\mathrm{\φ}\left(x_{j,k}\right))$

$s.t.\underset{i-11}{\overset{N}{\mathrm{\Σ}}}{y}_{i,k}{\mathrm{\α}}_{i,k}=0$

0≤α _i,k≤C,i=11,12,...,N,k=1,2,...,144

α _i,kAnd α _j,kThe mathematical model i that is used for the Support Vector Machines Optimized parameter and j Lagrange multiplier that expression is constantly corresponding with k.

The SVM prediction model constantly corresponding with k is:

$f\left(x_k\right)=\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_k\right)\·\mathrm{\φ}\left(x_{i,k}\right))+y_{j,k}-\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_{i,k}\right)\·\mathrm{\φ}\left(x_{j,k}\right))$

Wherein, x _kThe input vector of expression forecast sample, N represents the collection number of days of historical data,

The optimal value of i the Lagrange multiplier that the expression basis mathematical model optimizing that be used for Support Vector Machines Optimized parameter corresponding with the k moment obtains, x _j,kConcentrate for k training sample constantly, be positioned at the input vector of support vector position j, y _j,kFor k training sample is constantly concentrated, be positioned at the target output value of support vector position j, φ () is illustrated under input and the inseparable condition of output linearity, and lower dimensional space is to the mapping of higher dimensional space, φ (x) can not occur separately, but with inner product φ (x _i,k) φ (x _j,k) form occur,, therefore model does not need to know the concrete form of φ (x), only need know its inner product φ (x _i,k) φ (x _j,k) form get final product, note φ (x _i,k) φ (x _j,k)=K (x _i,k, x _j,k), φ (x _k) φ (x _i,k)=K (x _k, x _i,k), K (x _i,k, x _j,k) and K (x _k, x _i,k) expressing support for the kernel function that vector machine adopts, C represents penalty factor.

Due to the high dimension (16 dimension) of input sample, described kernel function adopts gaussian kernel function, and for example concrete form is K (x _i,k, x _j,k)=exp (|| x _i,k-x _j,k|| ²/ 2 σ ²).

Compare with the method for tradition application electric automobile during traveling Predicting Performance Characteristics charging station load, the present invention adopts the SVM forecast model based on pretreated historical data can effectively improve the precision of electric automobile charging station load prediction, and error reduces level such as table 1.By finding out in table, all have clear improvement on mean absolute error and two indexs of root-mean-square error based on the more traditional method based on the Vehicle Driving Cycle characteristic of the charging station load prediction of SVM model: mean absolute error reduces by 69.9%, and root-mean-square error reduces by 64.5%.The charging station load sequence of two kinds of method predictions and actual charging station load sequence such as Fig. 2, curve in figure shows, charging station based on SVM forecast model load prediction curve of the present invention is substantially identical with actual charging station load curve tendency, and predicted value and actual value approach; And based on the charging station load prediction curve of electric automobile during traveling characteristic and the peak value time of occurrence of actual charging station load curve, differ larger, and between its predicted value and actual value, error is more obvious.

Table 1 is based on electric automobile during traveling Predicting Performance Characteristics and SVM forecast result of model

The present invention, take certain electric automobile charging station as example, illustrates that this invention is in the validity that improves electric automobile charging station load prediction precision.Based on electric automobile during traveling characteristic and SVM model prediction electric automobile charging station next day load sequence, the error level of two kinds of methods is as shown in table 1 respectively, predicts the outcome and charges the load actual value as shown in Figure 2.Result shows, based on the charging station load prediction of SVM model,, being significantly improved than the prediction based on ride characteristic aspect precision of prediction, all has clear improvement on mean absolute error and two indexs of root-mean-square error.

Travelling of electric automobile has very large randomness, and its time and battery dump energy to the charging station charging also has very large randomness, causes the method precision of traditional application electric automobile during traveling Predicting Performance Characteristics charging station load lower.Promote the growth of time along with electric automobile, the historical informations such as the charging load in the charging station database, electric automobile quantity, charging set quantity are very sufficient.The present invention is directed to deficiency and problem that prior art exists, application support vector machine (Support Vector Machines, SVM) method, consider each factor that affects electric automobile charging station, comprise historical charging load, all attributes, red-letter day attribute, temperature, weather condition, the vehicle number of operation, the charging pile number of normal operation etc., the prediction charging station is loaded.The method can effectively improve the precision of electric automobile charging station load prediction, has simultaneously stronger versatility and engineering adaptability.

The present invention is docked with electric automobile charging station database (iNova database), can realize the prediction of charging station load.Utilize method of the present invention can effectively improve the precision of electric automobile charging station load prediction, improve adaptivity and the engineering practicability of electric automobile charging station load forecasting method, for power scheduling department adjusts operation plan, adjustment System margin capacity, optimizes genset and exert oneself etc. more reliable foundation is provided.The present invention can output in database predicting the outcome in addition, and the form with curve represents simultaneously, facilitates the secondary treating of data and intuitively represents.

Claims (9)

1. load forecasting method of the electric automobile charging station based on support vector machine, it is characterized in that: this Forecasting Methodology comprises the following steps:

1) charging station load and the historical data that affect factor of charging station load before Collection and Forecast Day, described factor comprise all attributes, red-letter day attribute, temperature, weather condition, charging station service vehicle number and the charging set number that works;

2) each affect the value of the factor of charging station load to determine prediction day, and wherein, temperature and weather condition determine according to weather forecast, and the several data that provide according to charging station of charging station service vehicle number and normal operation charging set are definite;

3) to historical data and step 2) value determined makes normalized;

4) after step 3), historical data and step 2 that utilization and prediction day k are constantly corresponding) definite value structure training sample set and forecast sample collection;

5) utilize the training sample set training to obtain the Lagrange multiplier of the support vector machine regression function constantly corresponding with k, according to described Lagrange multiplier, set up the SVM prediction model constantly corresponding with k;

6) forecast sample collection substitution SVM prediction model is obtained the charging station load prediction value constantly corresponding with k;

7) repeating step 4), to step 6), obtain predicting a day charging station load prediction value corresponding to each moment.

2. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 1, it is characterized in that: described all attributes are designated as W, W ∈ { 1,2,3,4,5,6,7}, W=1 corresponding Monday, W=2 corresponding Tu., W=3 corresponding Wednesday, W=4 corresponding Thursday, W=5 corresponding Friday, W=6 corresponding Saturday, W=7 corresponding Sun.; Red-letter day, attribute was designated as F, and { 0,1}, if get F=1 red-letter day, otherwise get F=0 to F ∈; Weather condition is designated as A, A ∈ 1,2,3}, and the A=1 correspondence is fine, and A=2 is corresponding cloudy, the corresponding sleet of A=3.

3. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 1 is characterized in that: described historical data for electric automobile charging station data-base recording in before prediction day 100-200 days about the charging station load and the data message that affect factor that charging station loads.

4. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 1, it is characterized in that: before historical data is made normalized, the historical data that gathers is scanned, when if in historical data, one day, charging station load constantly was for negative value, with modified value, replace negative value, modified value is the mean value of forward and backward two days of described one day corresponding charging station constantly load.

5. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 1, it is characterized in that: the construction method of described training sample set is: the front 10 days k charging station load value L={L constantly that gets i day _{I-1, k}, L _{I-2, k}, L _{I-3, k}, L _{I-4, k}, L _{I-5, k}, L _{I-6, k}, L _{I-7, k}, L _{I-8, k}, L _{I-9, k}, L _{I-10, k}, all attribute W of i day _i, i day attribute F in red-letter day _i, i day temperature T _i, i day weather condition A _i, i day charging station service vehicle counts B _iAnd the charging set that works i day is counted C _iInput vector as i day k training sample constantly:

x _i,k={L _i-1,k,L _i-2,k,L _i-3,k,L _i-4,k,L _i-5,k,L _i-6,k,L _i-7,k,L _i-8,k,L _i-9,k,L _i-10,k,W _i,F _i,T _i,A _i,B _i,C _i}

Get i day k charging station load L constantly _i,kTarget output value as training sample:

y _i,k=L _i,k

The training sample set of any time

I=11,12 ..., N, N are the collection number of days of historical data.

6. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 5, it is characterized in that: the construction method of described forecast sample collection is: the front 10 days k charging station load value L={L constantly that gets prediction day _{M-1, k}, L _{M-2, k}, L _{M-3, k}, L _{M-4, k}, L _{M-5, k}, L _{M-6, k}, L _{M-7, k}, L _{M-8, k}, L _{M-9, k}, L _{M-10, k}, the prediction day all attribute W _M, the prediction day attribute F in red-letter day _M, the prediction day temperature T _M, the prediction day weather condition A _M, a prediction day charging station service vehicle counts B _MAnd a prediction day normal operation charging set is counted C _MAs the input vector of k forecast sample constantly, M=N+1:

x _k={L _M-1,k,L _M-2,k,L _M-3,k,L _M-4,k,L _M-5,k,L _M-6,k,L _M-7,k,L _M-8,k,L _M-9,k,L _M-10,k,W _M,F _M,T _M,A _M,B _M,C _M}

。

7. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 6, it is characterized in that: the SVM prediction model constantly corresponding with k is:

$f\left(x_k\right)=\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_k\right)\·\mathrm{\φ}\left(x_{i,k}\right))+y_{j,k}-\underset{i=11}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i,k}^{*}{y}_{i,k}(\mathrm{\φ}\left(x_{i,k}\right)\·\mathrm{\φ}\left(x_{j,k}\right))$

Wherein, x _kThe input vector of expression forecast sample, N represents the collection number of days of historical data,

The optimal value that represents i Lagrange multiplier, x _j,kConcentrate for k training sample constantly, be positioned at the input vector of support vector position j, y _j,kConcentrate for k training sample constantly, be positioned at the target output value of support vector position j, φ (x _i,k) φ (x _j,k)=K (x _i,k, x _j,k), φ (x _k) φ (x _i,k)=K (x _k, x _i,k), K (x _i,k, x _j,k) and K (x _k, x _i,k) express support for the kernel function that vector machine adopts.

8. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 7, it is characterized in that: described kernel function adopts gaussian kernel function.

9. a kind of load forecasting method of electric automobile charging station based on support vector machine according to claim 1, it is characterized in that: described Forecasting Methodology is further comprising the steps of: call figure interface display instrument, the form of prediction day each charging station load prediction value constantly with curve represented.

Images