CN104834975A - Power network load factor prediction method based on intelligent algorithm optimization combination - Google Patents

Abstract

Disclosed is a power network load factor prediction method based on intelligent algorithm optimization combination. According to the method, first of all, based on such means of Granger examination and the like, numerous factors related to load factors are screened and left under the condition that no information is omitted, and the analysis accuracy of the load factor-related factors and the predication precision of the load factors are guaranteed. When the load factors are predicted, such intelligent algorithms of an RBF nerve network, a GRNN nerve network, an SVR algorithm and the like are taken into consideration, the advantages of less omitted information, less attention to internal relations, high prediction precision and the like of an intelligent algorithm are brought into full play compared to a conventional prediction method, and multiple prediction results are optimized and combined by use of a genetic algorithm so as to further improve the prediction precision. The method provided by the invention can be used for load factor precision of different time scopes such as year, month, date and the like, can also be applied to the load factor prediction of such classified users as large industrial users, residential users and the like, and provides certain theoretical support for research correlated with a load factor price.

Description

A kind of load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination

Technical field

The invention belongs to load rate of grid Forecasting Methodology technical field, particularly relate to a kind of load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination.

Background technology

Rate of load condensate is the index of power capacity producing level.The imagination of carrying out electricity price reform according to customer charge characteristic is just proposed in 2003 national " sales rate of electricity reform schemes ".Country in 2012 begins one's study the sales rate of electricity and rate of load condensate electricity price implementing method formulating and consider rate of load condensate factor, promotes power industry and national economy from extensive to the transition of Compact development with this.Cost factor is the basis of formulating electricity price, shares and calculate the electric cost of different user according to customer charge rate, is the key problem in technology that rate of load condensate electricity price is implemented.Based on this, the forecast analysis of rate of load condensate serves vital effect for the rational of rate of load condensate electricity price and popularization.

But the domestic and international research to load factor estimation is still less at present, but about the technical method relative maturity of load prediction.In load forecast, traditional load forecasting method has regressive prediction model, Random time sequence forecast model, grey forecasting model, expert system approach etc.The limitation of character and the various Forecasting Methodology itself such as consider the influenced many factors of rate of load condensate self, each factor action principle is failed to understand, traditional load forecasting method is difficult to the object reaching Accurate Prediction.Such as, regressive prediction model adopts the too simple linear model of structure to go to solve serious nonlinear problem, therefore cannot the various influence factors of detailed description rate of load condensate; Random time sequence forecast model process when modeling is complicated, considers also not comprehensive on the factor (as weather, economic dispatch) affecting rate of load condensate variation; The range of application of grey forecasting model is less, easily expands for long-term forecasting time error; The undue dependent Rule of expert system approach, universality is poor.

And some intelligent algorithms are as neural network, SVR, genetic algorithm etc., skip going into seriously the inherent action principle of rate of load condensate correlative factor, effectively can solve the mistake existed in traditional prediction method and simplify the problems such as process, omission influence factor, universality difference.And intelligent algorithm ubiquity algorithm parameter many foundations subjective experience is determined, lack the new problem of suitable theoretical direction.In order to head it off, need to reduce predicated error for target, rational optimal combination is carried out to predicting the outcome of relevant different artificial intelligence approach.

Summary of the invention

In order to solve the problem, the object of the present invention is to provide a kind of load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination.

In order to achieve the above object, the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination provided by the invention comprises the following step performed in order:

The first step: combing is carried out to the factor that may impact load rate of grid, data based on the data that these factors are corresponding, the timed sample sequence number of employing is not less than 10;

Second step: application Eviews software screens above many factors, rejects the factor with network load rate dependence difference;

3rd step: be normalized the corresponding basic data of sequence factor retained, to eliminate the impact of dimension on prediction;

4th step: select suitable Forecasting Methodology according to corresponding relation according to load rate of grid type from form below, and predict, to be predicted the outcome in conjunction with above-mentioned normalized related data based on different software platforms;

Table 1 rate of load condensate type is corresponding with Forecasting Methodology to be shown

5th step: through reverse reduction by the above-mentioned data be treated under normal dimension that predict the outcome;

6th step: using average absolute percentage error MAPE as fitness function, combined prediction mathematical model is utilized to be optimized combination, to obtain optimum prediction result predicting the outcome of RBF neural algorithm, GRNN neural network algorithm and SVR algorithm based on genetic algorithm.

In a first step, the described factor that may impact load rate of grid mainly comprises Factors Affecting Economic Development, dsm factor, temperature climatic factor, power grid environment factor and low-carbon economy development factors.

In second step, described application Eviews software screens above many factors, reject and with the method for the factor of network load rate dependence difference be: first carry out unit root test by the time series of Eviews software to above single factor, whether steady to consider each time series, if steadily, then do Granger causality analysis with load rate of grid; If not steady, then carry out co integration test with load rate of grid, if there is the whole relation of association, then do Granger causality analysis with load rate of grid further, otherwise this factor is cast out; By Granger causality analysis, if this factor be load rate of grid Granger because of, then retain this factor and predict for load rate of grid; Otherwise this factor is cast out.

In the third step, described normalization processing method is as follows: the basic data that the sequence factor retained above is relevant is normalized interval [0,1], if these data itself are namely interval [0,1] this step is then skipped in, as load rate of grid itself does not then need to process, formula is as follows:

$y=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(1\right)$

Wherein, x is numerical value before process, and y is numerical value after process, x _max, x _minbe respectively maximal value and the minimum value of this item number certificate.

In the 6th step, described combined prediction mathematical model is as follows:

${\hat{y}}_t=\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i{y}_{\mathrm{it}}(t=\mathrm{1,2},...,n),$ And $\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i=1---\left(2\right)$

Wherein, y _it(i=1,2 ..., k; T=1,2 ..., n) be the predicted value of i-th kind of Forecasting Methodology in t, w _iit is the weight coefficient of i-th kind of Forecasting Methodology.

The effect of the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination provided by the invention:

First the present invention leaves based on the screening when not drain message of the means such as Granger inspection factor, the accuracy of guaranteed load rate Study on Relative Factors and the precision of load factor estimation that a large amount of and rate of load condensate exists relation.When carrying out load factor estimation, consider the intelligent algorithms such as RBF neural, GRNN neural network, SVR, give full play to the advantage that intelligent algorithm drain message is few, do not go into seriously internal relations, the high relative traditional prediction method of precision of prediction, and use genetic algorithm to be optimized combination to improve precision of prediction further to multiple predicting the outcome.The method both can be used for the load factor estimation of the different time scopes such as year, month, day, can be used for again the load factor estimation of the sorted users such as large industrial user, resident, for the relevant research of rate of load condensate electricity price provides certain theory support.

Accompanying drawing explanation

Fig. 1 is the load rate of grid Forecasting Methodology process flow diagram based on intelligent algorithm optimal combination provided by the invention.

Fig. 2 is RBF genetic algorithm optimization result.

Fig. 3 is GRNN genetic algorithm optimization result.

Fig. 4 is that RBF, GRNN, SVM predict the outcome comparison diagram.

Fig. 5 is RBF, GRNN, SVM predicated error comparison diagram.

Fig. 6 is genetic algorithm optimization procedure chart.

Embodiment

Below in conjunction with the drawings and specific embodiments, the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination provided by the invention is described in detail.

First the present invention selects the factor relevant to rate of load condensate based on unit root test, cointegrating analysis with Granger test sieve, and then undertaken predicting by multi-intelligence algorithm and be optimized combination in conjunction with genetic algorithm to predicting the outcome, the Software tool related to mainly comprises EXCEL, Eviews, SPSS and Matlab.

As shown in Figure 1, the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination provided by the invention comprises the following step performed in order:

The first step: combing is carried out to the factor that may impact load rate of grid, data based on the data that these factors are corresponding, the timed sample sequence number of employing is not less than 10; Wherein, the factor that may impact load rate of grid mainly comprises Factors Affecting Economic Development, dsm factor, temperature climatic factor, power grid environment factor and low-carbon economy development factors etc.Causing result of calculation misalignment to prevent omitting information of forecasting, needing the various basic datas that extensive collection in a large number may be relevant to load rate of grid.

Second step: application Eviews software screens above many factors, rejects the factor with network load rate dependence difference; First unit root test is carried out by the time series of Eviews software to above single factor, whether steady to consider each time series, if steadily, then do Granger causality analysis with load rate of grid; If not steady, then carry out co integration test with load rate of grid, if there is the whole relation of association, then do Granger causality analysis with load rate of grid further, otherwise this factor is cast out; By Granger causality analysis, if this factor be load rate of grid Granger because of, then retain this factor and predict for load rate of grid; Otherwise this factor is cast out.

3rd step: be normalized the corresponding basic data of sequence factor retained, to eliminate the impact of dimension on prediction; The basic data that the sequence factor retained above is relevant be normalized interval [0,1], if namely these data itself then skip this step in interval [0,1], as load rate of grid itself does not then need to process, formula is as follows:

$y=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(1\right)$

Wherein, x is numerical value before process, and y is numerical value after process, x _max, x _minbe respectively maximal value and the minimum value of this item number certificate.

4th step: select suitable Forecasting Methodology according to corresponding relation according to load rate of grid type from form below, and predict, to be predicted the outcome in conjunction with above-mentioned normalized related data based on different software platforms;

Table 1 rate of load condensate type is corresponding with Forecasting Methodology to be shown

5th step: through reverse reduction by the above-mentioned data be treated under normal dimension that predict the outcome;

6th step: using average absolute percentage error MAPE as fitness function, combined prediction mathematical model is utilized to be optimized combination, to obtain optimum prediction result predicting the outcome of RBF neural algorithm, GRNN neural network algorithm and SVR algorithm based on genetic algorithm.

Due to the incomprehensiveness to future, single Forecasting Methodology often has very large risk, method larger for predicated error is given up blindly and may lose a part of information.In order to strengthen forecasting reliability, above RBF neural algorithm, GRNN neural network algorithm, SVR algorithm suitably can be combined, being fully utilized the information that various method provides, thus effectively be improved precision of prediction and reliability.Combined prediction mathematical model is as follows:

${\hat{y}}_t=\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i{y}_{\mathrm{it}}(t=\mathrm{1,2},...,n),$ And $\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i=1---\left(2\right)$

Wherein, y _it(i=1,2 ..., k; T=1,2 ..., n) be the predicted value of i-th kind of Forecasting Methodology in t, w _iit is the weight coefficient of i-th kind of Forecasting Methodology.

The essence of this method is that the optimization of each forecast model weight is determined, finds the weight that difference predicts the outcome, reduces error further.

Concrete case: this example is data based on the relevant historical data of 2000-2011 of Tianjin, and predict the yearly load factor of network system in 2012 in conjunction with RBF neural algorithm, GRNN neural network algorithm, SVR and genetic algorithm.

The first step, arranges the correlative factor that may impact yearly load factor, wherein involved factor and the historical data of correspondence as shown in the table:

Table 2 yearly load factor fundamentals of forecasting tables of data

Second step, application Eviews software screens above many factors, rejects the irrelevant factor poor with yearly load factor correlativity.

Unit root test and co integration test are done to each factor above, Granger Causality Test is done further to qualified factor.Analyze for Tianjin year GDP, unit root test result is obtained: the value of Prob is greater than 0.05 by Eviews software, and the critical value under 1%, 5%, 10% level is less than the value of the Augmented Dickey-Fuller of hypothesis, explanation can not refuse null hypothesis, and illustrates that the time series of Tianjin year GDP is not steady.

And then, Tianjin year GDP and Tianjin yearly load factor are carried out co integration test as correlated variables, can find, at least there are 2 whole relations of association, and the value of Prob is less than 0.05, therefore can carry out Granger Causality Test.

Can find out according to Tianjin year GDP and yearly load factor Granger Causality Test result, the value 0.0405 of first Prob is less than 0.05, then illustrate the annual GDP in Tianjin be not Tianjin yearly load factor Granger because of.

Successively above analysis is done to other factors such as GDP per capita, Different Industries structure proportion, population, finally find to can be considered the Granger of Tianjin yearly load factor because of factor comprise: primary industry structure proportion, per capita disposable income, the annual average price of steam coal.

3rd step, carries out pre-service to the basic data of above correlative factor.

Above basic data be normalized interval [0,1], formula is as follows:

$y=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(3\right)$

Wherein, x is numerical value before process, and y is numerical value after process, x _max, x _minbe respectively maximal value and the minimum value of this item number certificate.

Basic data after being normalized is as shown in the table:

The basic data of table 3 after screening and normalized

4th step, adopts RBF neural algorithm, GRNN neural network algorithm and SVR algorithm to predict yearly load factor according to the corresponding situation of table 1 respectively.

(1) predict based on RBF neural algorithm

Using the data of 2000-2008 years in upper table as training sample, the data of 2009-2012 years, as test samples, are predicted by RBF neural.In order to reduce predicated error, with the MAPE predicted the outcome for fitness function, the distribution density function Spr of genetic algorithm to RBF neural is adopted to be optimized.

Following table is the important parameter adopting genetic algorithm optimization, and unlisted parameter all gets default value:

Table 4RBF genetic algorithm optimization parameter

Parameter name	Parameter value
		The upper limit	1
Lower limit	0
		Algebraically	50
Population quantity	100

Fig. 2 is the final optimization pass result of genetic algorithm, can be found out by this figure, and along with not very increasing of algebraically, the value of MAPE constantly reduces, and optimizes stop in the 50th generation.When Spr=0.084 (0.083849), MAPE obtains optimal value.Spr=0.084 is substituted into neural network predict, the yearly load factor value tentatively obtaining 2009-2012 years is respectively 0.6231,0.6233,0.6231,0.6231.

(2) predict based on GRNN neural network algorithm

In order to reduce predicated error, with the MAPE predicted the outcome for fitness function, the smoothing factor of genetic algorithm to GRNN neural network is adopted to be optimized.

Following table is the important parameter adopting genetic algorithm optimization, and unlisted parameter all gets default value:

Table 5GRNN genetic algorithm optimization parameter

Parameter name	Value
		The upper limit	1
Lower limit	0
		Algebraically	50
Population quantity	100

Fig. 3 is the final optimization pass result of genetic algorithm.Can be found out by this figure, along with not very increasing of algebraically, the value of MAPE constantly reduces, and optimizes stop in the 50th generation.When Spr=0.930 (0.929996), MAPE obtains optimal value.Spr=0.930 is substituted into neural network predict, the yearly load factor value tentatively obtaining 2009-2012 years is respectively 0.5653,0.5789,0.5861,0.5832.

(3) predict based on SVR algorithm

In order to reduce predicated error, different parameter values being set and result is compared.

Wherein adopt SVM prediction, the yearly load factor value tentatively obtaining 2009-2012 years is respectively 0.5422,0.5458,0.5483,0.5492.

5th step, through inverse normalization by the above-mentioned data be treated under normal dimension that predict the outcome, as shown in the table:

Table 6 to predict the outcome contrast against the difference after normalization

6th step, using average absolute percentage error MAPE as fitness function, is optimized combination based on genetic algorithm to predicting the outcome under different Forecasting Methodology, finds the weight that difference predicts the outcome, reduce error further.

The method of the RBF network of genetic algorithm optimization, the GRNN network of genetic algorithm optimization and support vector machine is used to predict the yearly load factor of 2009-2012 respectively above.Forecasting Methodology dissimilar above, respectively based on different theories, provides different information, precision also difference to some extent of its prediction.Fig. 4, Fig. 5 are respectively RBF, GRNN, SVM and predict the outcome comparison diagram and RBF, GRNN, SVM predicated error comparison diagram.

Undertaken based on the data that load prediction obtains, by respective weights w by the method for the GRNN network of the RBF network of genetic algorithm optimization, genetic algorithm optimization and support vector machine ₁, w ₂and w ₃as the variable of genetic algorithm optimization, encode in floating number mode, by the equality condition constraint matrix representation of weights.Writing with MAPE is the fitness function mix of index,

Fig. 6 is genetic algorithm optimization process.Execution result finds, works as w ₁=0, w ₂=1 and w ₃when=0, MAPE is minimum is 2.1556.

It is below the contrast situation of the average absolute percentage error MAPE after RBF, GRNN, SVM and GA optimal combination, can find out that the predicated error after GA optimizes is identical with the predicated error of GRNN algorithm from figure below, in this example, the precision of prediction of GRNN algorithm is apparently higher than RBF and SVM algorithm, after GA combination, therefore only adopt the result of GRNN algorithm.

First load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination provided by the invention carries out data preparation, and to the root inspection of basic data executable unit, cointegrating analysis and Granger inspection, both the factor that a large amount of and load rate of grid exists relation can have been filtered out, will precision of prediction be caused to decline by drain message again, for the correlated influencing factors of load rate of grid and prediction lay the first stone.And then, obtain predicted value by neural network, support vector machine (SVR) scheduling algorithm and be optimized combination to predicting the outcome, the combination forecasting method that this precision is higher both may be used for the prediction of the rate of load condensates such as year, month, day, can be used for again the prediction of sorted users rate of load condensate, the formulation for rate of load condensate electricity price provides certain foundation.

Claims (5)

1. based on a load rate of grid Forecasting Methodology for intelligent algorithm optimal combination, it is characterized in that: it comprises the following step performed in order:

The first step: combing is carried out to the factor that may impact load rate of grid, data based on the data that these factors are corresponding, the timed sample sequence number of employing is not less than 10;

Second step: application Eviews software screens above many factors, rejects the factor with network load rate dependence difference;

3rd step: be normalized the corresponding basic data of sequence factor retained, to eliminate the impact of dimension on prediction;

4th step: select suitable Forecasting Methodology according to corresponding relation according to load rate of grid type from form below, and predict, to be predicted the outcome in conjunction with above-mentioned normalized related data based on different software platforms;

Table 1 rate of load condensate type is corresponding with Forecasting Methodology to be shown

5th step: through reverse reduction by the above-mentioned data be treated under normal dimension that predict the outcome;

6th step: using average absolute percentage error MAPE as fitness function, combined prediction mathematical model is utilized to be optimized combination, to obtain optimum prediction result predicting the outcome of RBF neural algorithm, GRNN neural network algorithm and SVR algorithm based on genetic algorithm.

2. the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination according to claim 1, it is characterized in that: in a first step, the described factor that may impact load rate of grid mainly comprises Factors Affecting Economic Development, dsm factor, temperature climatic factor, power grid environment factor and low-carbon economy development factors.

3. the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination according to claim 1, it is characterized in that: in second step, described application Eviews software screens above many factors, reject and with the method for the factor of network load rate dependence difference be: first carry out unit root test by the time series of Eviews software to above single factor, whether steady to consider each time series, if steadily, then do Granger causality analysis with load rate of grid; If not steady, then carry out co integration test with load rate of grid, if there is the whole relation of association, then do Granger causality analysis with load rate of grid further, otherwise this factor is cast out; By Granger causality analysis, if this factor be load rate of grid Granger because of, then retain this factor and predict for load rate of grid; Otherwise this factor is cast out.

4. the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination according to claim 1, it is characterized in that: in the third step, described normalization processing method is as follows: the basic data that the sequence factor retained above is relevant is normalized interval [0,1], if these data itself are namely interval [0,1] then skip this step in, as load rate of grid itself does not then need to process, formula is as follows:

$y=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(1\right)$

Wherein, x is numerical value before process, and y is numerical value after process, x _max, x _minbe respectively maximal value and the minimum value of this item number certificate.

5. the load rate of grid Forecasting Methodology based on intelligent algorithm optimal combination according to claim 1, it is characterized in that: in the 6th step, described combined prediction mathematical model is as follows:

${\hat{y}}_t=\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i{y}_{\mathrm{it}}(t=\mathrm{1,2},...,n),$ And $\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{w}_i=1---\left(2\right)$

Wherein, y _it(i=1,2 ..., k; T=1,2 ..., n) be the predicted value of i-th kind of Forecasting Methodology in t, w _iit is the weight coefficient of i-th kind of Forecasting Methodology.