CN1731446A - Comprehensive load prediction method based on minimum information loss - Google Patents

Abstract

The invention relates to a synthetic burden forecasting method based on the minimum information loss in the field of electrical power system burden forecasting technology. The method comprises: choosing a history reference day, doing virtual forecasting, doing the communication channel parameter estimation and the communication source parameter estimation of the M kinds of algorithm to the time t by the virtual forecasting result and the real time burden value of the history reference day, doing single algorithm forecasting, establishing the minimum loss object function by the communication channel and the communication source parameter estimation and the single algorithm forecasting result, computing the object function to obtain the forecasting result vt of the forecasting day t time, repeating the above steps to obtain the burden forecasting sequence of the entire day T point.

Description

Comprehensive load prediction method based on minimum information loss

Technical field

The invention belongs to power system load prediction (power system load forecasting) technical field, particularly the integrated forecasting method in the load prediction (combined forecasting method).

Background technology

The power system load prediction is meant from known economy, social development and electric system conditions of demand, by analysis and research to historical data, explore the inner link and the rule of development between the things, with following period predicting the outcome of correlative factor such as economy, social development situation be foundation, electricity needs is made in advance estimation and prediction, be requisite basic link in Power System Planning, operation and the marketing, it predicts the outcome and is used for power source planning, Electric Power Network Planning, generation schedule, the combination of unit economy, transaction plan, electricity consumption plan etc.Load Forecasting result accurately whether, security, reliability and the economy of Operation of Electric Systems all had significant impact.

In the power system load prediction, single load prediction algorithm all is to set up according to certain specific Changing Pattern of historical load data, for the electric load that meets this kind Changing Pattern good prediction effect is arranged.Yet the rule of development of electric load is quite complicated, be difficult to predict with single algorithm, predicting the outcome of many prediction algorithms organically must be combined, could form the description properer or more complete, integrated forecasting method that Here it is the electric load rule of development.The integrated forecasting method can organically make up various algorithms, can effectively improve precision of prediction, is extensively adopted by power industry circle.

At present, the integrated forecasting method of electric load mainly contains two big classes.

The first kind is determined algorithm weights by the prediction effect of estimating each algorithm.For example: adopt index that each algorithm is made fuzzy evaluation, again by comparing the comprehensive weight of each algorithm of quality relation calculating between each fuzzy set.At present, the weight calculation of this class integrated approach lacks the criterion of generally acknowledging.

Second class, by the match of historical data being determined the weight of each algorithm in the integrated forecasting method, target is to make the virtual prognostication match residual sum minimum between sequence and the actual history load sequence as a result after the weighting, and this class integrated forecasting method has obtained widespread use in practice.

Mo Weiren, Zhang Baiming, Sun Hongbin, beard weighing apparatus are in " discussion of short-term load integrated forecasting model " (Automation of Electric Systems, in January, 2004, the 28th the volume, the 1st phase, 30-34 page or leaf) in the mechanism of the above-mentioned second class integrated forecasting method is studied and is inquired into.It realizes that principle is: seek out and the prediction day similar historical date (being historical reference day) of various influence factors; Adopt different prediction algorithms that it is carried out virtual prognostication; Relatively this day actual history load with virtual prognostication result's match accuracy to determine the weight allocation of various algorithms; The weight of using resulting each algorithm is predicted the load prediction of day.Its concrete implementation step is as follows:

1. choose historical reference day: take all factors into consideration various influence factors, be prediction day N historical reference day of searching.If the load sampling number of every day is T (the value relative fixed is generally 96 or 288 in practice), then the actual history of each historical reference day load sequence can be expressed as:

(x _n1，x _n2，…，x _nT)，n＝1，2，…，N

Wherein, x _NtRepresent the Japan-China t of n historical reference actual load value constantly, t=1,2 ..., T;

2. virtual prognostication: adopt M kind prediction algorithm, each historical reference daily load is carried out virtual prognostication, the virtual prognostication that obtains each historical reference day is sequence as a result:

( _n1m， _n2m，…， _nTm)，n＝1，2，…，N，m＝1，2，…，M

Wherein,  _NtmRepresent the predicted value that m kind algorithm is loaded constantly to the Japan-China t of n historical reference, t=1,2 ..., T;

3. optimal weights calculates: actual history load and corresponding virtual according to each historical reference day predict the outcome, and set up mathematical modulo pattern (1):

$\left\{\begin{array}{c}\min E=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{r}_n^2\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{(x_{\mathrm{nt}}-\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{w}_m{\hat{x}}_{\mathrm{ntm}})}^2\\ \underset{m=1}{\overset{M}{\mathrm{\Σ}}}{w}_m=1\end{array}\right.---\left(1\right)$

Find the solution this mathematical model, obtain and each algorithm optimal weights w one to one _m(m=1,2 ..., M).w _mCorresponding m kind algorithm; (1) in the formula, parameter r _n ²Be weight coefficient, it is determined according to the principle of " near big and far smaller " in advance according to each historical reference day and the prediction degree of closeness of day; From prediction day near more pairing r of historical reference day _n ²Be worth greatly more, purpose is the historical reference daily load that makes the preferential match of this Forecasting Methodology nearest;

4. single algorithm predicts: adopt M kind algorithm that prediction day is predicted, obtain predicting the outcome of each single algorithm:

(z _1m，z _2m，…，z _Tm)，m＝1，2，…，M

Wherein, z _TmRepresent the predicted value of m kind algorithm to prediction day t sampled point load;

5. weighted sum: the predicting the outcome of M kind algorithm that obtains in the step 4 is weighted summation, calculates the final z that predicts the outcome _t, that is:

$z_t=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\left(w_m{z}_{\mathrm{tm}}\right),t=\mathrm{1,2},\·\·\·,T---\left(2\right)$

(2) in the formula, weight w _mGet the optimal weights that obtains in the step 3.

Above-mentioned this class integrated forecasting method is based upon on the following hypothesis, if i.e.: integrated forecasting method must be got well the historical data match, then its accuracy of predicting is inevitable high.Yet the model that fitting precision is high might not just have good predicting the outcome.Particularly when the randomness of load variations was big, over-fitting (overfiting) phenomenon can appear in the integrated approach that fitting precision is high.At this moment, the weight combination the highest to the historical data fitting precision, its precision that predicts the outcome may not be desirable.

Summary of the invention

The objective of the invention is to propose a kind of new comprehensive load prediction method based on minimum information loss (MIL, minimum information loss) for overcoming the weak point of existing integrated forecasting method.The present invention is from the angle of information loss, can make full use of the information in the data such as historical load, thereby avoided the over-fitting to the historical reference sample, is more suitable for can improving the precision that predicts the outcome in the load prediction of the big electrical network of load randomness.

The comprehensive load prediction method based on minimum information loss (MIL) that the present invention proposes may further comprise the steps:

1) chooses historical reference day: take all factors into consideration various influence factors, for N historical reference day determined in a prediction day searching; If the load sampling number of every day is T, then the actual history of each historical reference day load sequence can be expressed as:

(x _n1，x _n2，…，x _nT)，n＝1，2，…，N

Wherein, x _NtThe actual load value of representing Japan-China t the sampled point of n historical reference, t=1,2 ..., T;

2) virtual prognostication: adopt M kind algorithm respectively the load of the t moment point among the load sampling number T of each historical reference day to be carried out virtual prognostication, obtain the result of virtual prognostication:

 _ntm，n＝1，2，…，N，m＝1，2，…，M

 _NtmRepresent the predicted value of m kind algorithm to Japan-China t the sampled point load of n historical reference;

3) parameter estimation:, this moment t is done the estimation of channel parameter of M kind algorithm and the estimation of information source parameter according to the result of virtual prognostication and the actual load value of historical reference day.

4) single algorithm predicts: adopt M kind algorithm respectively the t that predicts day to be loaded constantly and predict, obtain the z that predicts the outcome of each single algorithm respectively _Tm: z _TmRepresent the predicted value that m kind algorithm is loaded constantly to prediction day t;

5) integrated forecasting of single-point load: set up the objective function of information loss minimum according to step 3) and step 4) result of calculation, find the solution this objective function, obtain predicting day t v that predicts the outcome constantly _t

6) integrated forecasting of multiple spot load: for the moment point of T altogether of prediction day whole day (t value from 1 to T), repeat above step 2 respectively, just can obtain the load prediction value sequence of predicting that day whole day T is ordered: (v to step 5 ₁, v ₂..., v _T).

The channel parameter of moment t to be estimated can be the covariance matrix B of M kind algorithm channel in the above-mentioned step 3) _t, the information source parameter of moment t to be estimated can comprise information source average μ _tWith the information source variances sigma _St ²

Wherein, the covariance matrix B of the M kind algorithm channel of moment t _tEstimation formulas be:

In the formula, σ _Ctm ²Expression is the variance of the m kind algorithm channel of t constantly, Cov _Ct(i, j) i of expression moment t and the covariance of j kind algorithm interchannel adopt following formula to estimate respectively:

${\mathrm{\σ}}_{\mathrm{Ctm}}^2=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{({\hat{x}}_{\mathrm{ntm}}-x_{\mathrm{nt}})}^2,t=\mathrm{1,2},\·\·\·,T,m=\mathrm{1,2},\·\·\·,M---\left(4\right)$

${\mathrm{Cov}}_{\mathrm{Ct}}(i,j)=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}({\hat{x}}_{\mathrm{nti}}-x_{\mathrm{nt}})({\hat{x}}_{\mathrm{ntj}}-x_{\mathrm{nt}}),t=\mathrm{1,2},\·\·\·,T,i,j=\mathrm{1,2},\·\·\·,M$ And i ≠ j (5)

The information source average μ of moment t _tWith the information source variances sigma _St ²Adopt following formula to estimate respectively:

${\mathrm{\μ}}_t=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{x}_{\mathrm{nt}},t=\mathrm{1,2},\·\·\·,T---\left(6\right)$

${\mathrm{\σ}}_{\mathrm{St}}^2=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{(x_{\mathrm{nt}}-{\mathrm{\μ}}_t)}^2,t=\mathrm{1,2},\·\·\·,T---\left(7\right)$

The objective function of the information loss minimum in the above-mentioned step 5) can comprise:

When the channel Normal Distribution, information source is obeyed when evenly distributing, and the objective function of t is constantly:

$\min I_{\mathrm{loss}}(V_t;Z_t)=\frac{1}{2}{(Z_t-V_t)}^T{B}_t^{-1}(Z_t-V_t)---\left(8\right)$

In the formula, V _t=[v _t, v _t..., v _t] ^TBe the information source vector of the M dimension of t constantly, v _tBe estimated value to the load to be predicted of moment t, Z _t=[z _T1, z _T2..., z _Ti..., z _TM] ^TBe the stay of two nights vector of the M dimension of t constantly, z _TiBe the predicting the outcome of i kind load prediction algorithm of t constantly.I _Loss(V _tZ _t) overall information loss when expression is carried out comprehensive load prediction to moment t;

When the channel Normal Distribution, during the information source Normal Distribution, the objective function of t is constantly:

$\min I_{\mathrm{loss}}(V_t;Z_t)=\frac{1}{2}{(Z_t-V_t)}^T{B}_t^{-1}(Z_t-V_t)+{(v_t-{\mathrm{\μ}}_t)}^2\frac{1}{2{\mathrm{\σ}}_{\mathrm{St}}^2}.---\left(9\right)$

Inventive principle

In essence, the integrated forecasting method of load is a kind of decision-making technique of informix, and just the information about the prediction load that provides according to various single prediction algorithms is inferred the process of prediction load actual value.The result of integrated forecasting method approaches to predict the best estimate of load time of day most, according to information theory, makes that just information loss minimizes in the integrated approach forecasting process.

The present invention is on the information theory basis, at first set up the information source and the channel model of various load prediction algorithms: information source is meant real load value, channel is meant various load prediction algorithms, and the stay of two nights is meant predicting the outcome of load prediction algorithm, comprehensive load prediction method is according to the information that the stay of two nights provides information source to be made optimum estimate, just the estimation of minimum information loss.

In order to illustrate principle, below above-mentioned comprehensive load prediction method based on minimum information loss (MIL) is simply derived.

In the conventional information theory, information source is meant the sender of information, and channel is the media of information transmission, and the stay of two nights is the recipient of information.There is noise in the channel, can produces information and disturb.The notion of information source, channel and the stay of two nights in the information theory is introduced in the load prediction integrated approach, can set up the generalized channel model of comprehensive load prediction method: the actual value that predict electric load is an information source, various single load prediction algorithms are channels, and predicting the outcome of each algorithm output is the stay of two nights.The error that occurs in the forecasting process just is equivalent to the noise that exists in the channel, and the combined process of integrated forecasting method is exactly to remove the information process of reconstruction of noise, with the objective function of minimum information loss (MIL) as the integrated forecasting method, that is:

$\underset{V}{\min }{I}_{\mathrm{loss}}(V;Z)---\left(10\right)$

Wherein, Z represents predicting the outcome of various single prediction algorithms, and V represents the integrated forecasting result possible to this vector that predicts the outcome, I _LossThe overall information loss amount that produces in the expression integrated forecasting process.

Because generally exist correlativity respectively forming of integrated forecasting method between the algorithm.For example the temperature when prediction day in summer rises sharply, and minus deviation may appear in linear extrapolation and exponential smoothing predict the outcome simultaneously.Considering correlativity between such algorithm, comprehensively is a compound channel with multiple single prediction algorithm.Wherein, the input V=[v of channel, v ..., v] and be a M n dimensional vector n, v is the actual value of load to be predicted; The output Z=[z of channel ₁, z ₂..., z _i..., z _M] also be a M n dimensional vector n, z _iBe predicting the outcome of the single prediction algorithm of i kind, i=1,2 ..., M.Introduce F _CThe probability density function of expression channel I/O vector statistics relation can be expressed as F _C(V Z), is a binary vector bounded continuous function.

According to information theory, the information loss calculating formula of compound channel, information source is expressed as (11) (12) two formulas respectively:

$I_{\mathrm{loss},C}(V;Z)=\ln \frac{F_C\left(\tilde{V};Z\right)}{F_C(V;Z)}---\left(11\right)$

$I_{\mathrm{loss},S}\left(V\right)=\ln \frac{f_S\left(\underset{~}{v}\right)}{f_S\left(v\right)}---\left(12\right)$

Wherein, f _S(v) be the probability density function of the one-component v of information source V, Be to make f _S(v) obtain peaked information source component value. makes F _C(V Z) gets peaked information source vector value.

By (11) (12) two formulas, the overall information loss in the formula (10) is:

$I_{\mathrm{loss}}(V;Z)=I_{\mathrm{loss},C}(V;Z)+I_{\mathrm{loss},S}\left(V\right)=\ln \frac{F_C(\tilde{V};Z)}{F_C(V;Z)}+\ln \frac{f_S\left(\underset{~}{v}\right)}{f_S\left(v\right)}---\left(13\right)$

In actual applications, the I/O statistical relationship of channel is obeyed M dimension normal distribution (annotating: also can be other distribution) usually, and by formula (11), the channel information loss is:

$I_{\mathrm{loss},C}(V;Z)=\ln \frac{F_C(\tilde{V};Z)}{F_C(V;Z)}=\frac{1}{2}{(Z-V)}^T{B}^{-1}(Z-V)---\left(14\right)$

Wherein, B is the covariance matrix of algorithm channel.

In actual applications, information source is obeyed usually and is evenly distributed or normal distribution.

Distribute if information source is obeyed evenly, then by formula (12), the information source information loss is:

$I_{\mathrm{loss},S}\left(V\right)=\ln \frac{f_S\left(\underset{~}{v}\right)}{f_S\left(v\right)}=\mathrm{<figure-callout\; id="1"\; label="ln"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">ln</figure-callout>}1=0$

If information source Normal Distribution N (μ, σ _S ²), then by formula (12), the information source information loss is:

$I_{\mathrm{loss},S}\left(V\right)=\ln \frac{f_S\left(\underset{~}{v}\right)}{f_S\left(v\right)}={(v-\mathrm{\&mu;})}^2\frac{1}{2{\mathrm{\&sigma;}}_S^2}---\left(16\right)$

Can obtain two kinds of practical objective functions based on the comprehensive load prediction method of MIL:

1, information source is obeyed evenly and is distributed, the channel Normal Distribution, and objective function is:

$\min I_{\mathrm{loss}}(V;Z)=\frac{1}{2}{(Z-V)}^T{B}^{-1}(Z-V)---\left(17\right)$

2, the equal Normal Distribution of information source and channel, objective function is:

${\min I}_{\mathrm{loss}}(V;Z)=\frac{1}{2}{(Z-V)}^T{B}^{-1}(Z-V)+{(v-\mathrm{\&mu;})}^2\frac{1}{2{\mathrm{\&sigma;}}_S^2}---\left(18\right)$

Parameter in formula (17) and (18) is all replaced to t parameter constantly, and then formula (17) and (18) are formula (8) and (9).

Technical characterstic and effect

Comprehensive load prediction method of the present invention is a kind of integrated forecasting method based on minimum information loss (MIL, Minimum informationloss).The innovative point of this method mainly is to understand the load prediction algorithm with the viewpoint of communication theory, the channel model of load prediction algorithm is proposed, and on the basis of load prediction algorithm channel model, set up the objective function of minimum information loss, utilize this objective function can obtain the optimum estimate of actual load value.

Method that the present invention adopts and existing methods significantly difference have: (1) the present invention adopts minimum information loss as objective function, scientifically predicting the outcome of each single prediction algorithm carried out comprehensively, owing to be not that direct historical data fitting precision to the integrated forecasting method claims, thereby avoided the problem of over-fitting, can effectively improve the big accuracy of forecasting of randomness.(2) among the present invention the objective function by minimum information loss find the solution the net result that directly obtains load prediction, cancelled the notion of algorithm weights in the existing method, this and existing method make up the thinking difference that obtains load prediction results according to weight again by the computational algorithm weight.(3) in the integrated forecasting method that the present invention proposes, consider the correlativity between the various single prediction algorithms, also can take into account the statistical property that load distributes.

Description of drawings

The step block diagram that Fig. 1 proposes for the present invention based on the comprehensive load prediction method of MIL

Embodiment

The comprehensive load prediction method based on minimum information loss that the present invention proposes reaches accompanying drawing in conjunction with the embodiments and is described in detail as follows,

The general steps of the inventive method comprises as shown in Figure 1:

Step 1, choose historical reference day: taking all factors into consideration various influence factors, is that prediction day is sought and to be determined N historical reference day; If the load sampling number of every day is T;

Step 2, virtual prognostication: adopt M kind algorithm respectively the load of the t moment point among the load sampling number T of each historical reference day to be carried out virtual prognostication;

Step 3, parameter estimation:, this moment t is done the estimation of channel parameter of M kind algorithm and the estimation of information source parameter according to the result of virtual prognostication and the actual load value of historical reference day;

Step 4, single algorithm predicts: adopt M kind algorithm respectively the t that predicts day to be loaded constantly and predict, obtain predicting the outcome of each single algorithm respectively;

The integrated forecasting of step 5, single-point load:, obtain predicting predicting the outcome that day t loads constantly by finding the solution the objective function of minimum information loss;

The integrated forecasting of step 6, multiple spot load: for the moment point of T altogether of prediction day whole day (t value from 1 to T), repeat above step 2 respectively, just can obtain the load prediction value of predicting that day whole day T is ordered to step 5.

Short-term load forecasting process with certain actual electric network illustrates the specific embodiment of the present invention as embodiment.The prediction day of present embodiment is on November 17th, 2004, establishes to have 96 future positions (T=96) in one day, and per 15 minutes is a future position, and the load actual value of prediction day 0: 15 (t=1) is 3376MW.In the integrated forecasting method, have following four kinds (M=4) single load prediction algorithm to be adopted: linear extrapolation, exponential smoothing, all peak value predicted methods and electric weight trend-based forecasting, these algorithms all are existing ripe algorithms.

Step 1, choose historical reference day: prediction day (on November 17th, 2004) is Wednesday, belongs to working day.Therefore select 7 working days before November 17 as historical reference day (N=7), be respectively: November 5, November 8 to November 12, November 15.The actual load of 0: 15 each historical reference day is expressed as x _N1, wherein, subscript n=1,2 ..., 7 corresponding 7 historical reference days.

Step 2, virtual prognostication: adopt aforementioned four kinds of (M=4) single prediction algorithms, the load of 0: 15 each historical reference day is carried out virtual prognostication.The virtual prognostication result who obtains is expressed as  _N1m, subscript n=1,2 wherein ..., 7 corresponding 7 historical reference days, subscript m=1,2,3,4 corresponding 4 kinds of single load prediction algorithms.

Step 3, parameter estimation: the result according to virtual prognostication makes estimation to channel and information source parameter.

The channel errors of various algorithms is obeyed the higher-dimension normal distribution usually in the reality, and the variance of m kind algorithm channel adopts following formula to estimate:

${\mathrm{\&sigma;}}_{C1m}^2=\frac{1}{7}\underset{n=1}{\overset{7}{\mathrm{\&Sigma;}}}{({\hat{x}}_{n1m}-x_{n1})}^2---\left(19\right)$

Wherein, m=1,2,3,4.

The covariance of i and j kind algorithm interchannel adopts following formula to estimate:

$\mathrm{Co}{v}_{C1}(i,j)=\frac{1}{7}\underset{n=1}{\overset{7}{\mathrm{\&Sigma;}}}({\hat{x}}_{n1i}-x_{n1})({\hat{x}}_{n1j}-x_{n1})---\left(20\right)$

I wherein, j=1,2,3,4 and i ≠ j.

Obtain the covariance matrix formula of 4 kinds of algorithm channels:

${\mathrm{<figure-callout\; id="1"\; label="B"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">B</figure-callout>}}_1=\left[\begin{array}{cccc}{\mathrm{\&sigma;}}_{C11}^2& {\mathrm{Cov}}_{C1}\left(\mathrm{1,2}\right)& {\mathrm{Cov}}_{C1}\left(\mathrm{1,3}\right)& {\mathrm{Cov}}_{C1}\left(\mathrm{1,4}\right)\\ {\mathrm{Cov}}_{C1}\left(\mathrm{2,1}\right)& {\mathrm{\&sigma;}}_{C12}^2& {\mathrm{Cov}}_{C1}\left(\mathrm{2,3}\right)& {\mathrm{Cov}}_{C1}\left(\mathrm{2,4}\right)\\ {\mathrm{Cov}}_{C1}\left(\mathrm{3,1}\right)& {\mathrm{Cov}}_{C1}\left(\mathrm{3,2}\right)& {\mathrm{\&sigma;}}_{C13}^2& {\mathrm{Cov}}_{C1}\left(\mathrm{3,4}\right)\\ {\mathrm{Cov}}_{C1}\left(\mathrm{4,1}\right)& {\mathrm{Cov}}_{C1}(4,2)& {\mathrm{Cov}}_{C1}\left(\mathrm{4,3}\right)& {\mathrm{\&sigma;}}_{C14}^2\end{array}\right]---\left(21\right)$

In the present embodiment, the result of calculation of covariance matrix is:

${\mathrm{<figure-callout\; id="1"\; label="B"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">B</figure-callout>}}_1=\left[\begin{array}{cccc}234042.3& -9524.6& 9124.9& 1595.7\\ -9524.6& 84767.8& -12990.5& -1070.6\\ 9124.9& -12990.5& 20334.6& 2813.1\\ 1595.7& -1070.6& 2813.1& 15996.7\end{array}\right]$

In the present embodiment, load information source Normal Distribution, following formula is adopted in the estimation of information source average:

${\mathrm{\&mu;}}_1=\frac{1}{7}\underset{n=1}{\overset{7}{\mathrm{\&Sigma;}}}{x}_{n1}---\left(22\right)$

Following formula is adopted in the estimation of information source variance:

${\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">S</figure-callout>}1}^2=\frac{1}{7}\underset{n=1}{\overset{7}{\mathrm{\&Sigma;}}}{(x_{n1}-{\mathrm{\&mu;}}_1)}^2---\left(23\right)$

Concrete computation process is omited, and result of calculation is as follows: the average of information source is μ ₁=3403, variance is ${\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">S</figure-callout>}1}^2=2245.3.$

Step 4, single algorithm predicts: adopt four kinds of single prediction algorithms that the load of prediction day 15 minutes (t=1) is predicted that predicting the outcome is respectively at 0 o'clock:

Linear extrapolation: 3466MW;

Exponential smoothing: 3243MW;

All peak value predicted methods: 3541MW;

Electric weight trend-based forecasting: 3379MW.

That is: Z ₁=(3466,3243,3541,3379) ^T

The integrated forecasting of step 5, single-point load: in the present embodiment, the equal Normal Distribution of information source and channel based on the objective function of the integrated forecasting method of MIL is:

$\underset{v_t}{\min }{I}_{\mathrm{loss}}(V_t,Z_t)=\frac{1}{2}{(Z_t-V_t)}^T{B}_t^{-1}(Z_t-V_t)-{(v_t-{\mathrm{\&mu;}}_t)}^2\frac{1}{2{\mathrm{\&sigma;}}_{\mathrm{St}}^2}---\left(24\right)$

In the present embodiment, objective function is

$\underset{v_1}{\min }{I}_{\mathrm{loss}}({\mathrm{<figure-callout\; id="1"\; label="V"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">V</figure-callout>}}_1;Z_1)={\frac{1}{2}\left[\begin{array}{c}3466-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3243-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3541-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3379-v_1\end{array}\right]}^T\left[\begin{array}{cccc}234042.3& -9524.6& 9124.9& 1595.7\\ -9524.6& 84767.8& -12990.5& -1070.6\\ 9124.9& -12990.5& 20334.6& 2813.1\\ 1595.7& -1070.6& 2813.1& 15996.7\end{array}\right]\left[\begin{array}{c}3466-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3243-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3541-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\\ 3379-{\mathrm{<figure-callout\; id="1"\; label="v"\; filenames="A20051010240200121.png"\; state="\{\{state\}\}">v</figure-callout>}}_1\end{array}\right]+\frac{{(v_1-3403)}^2}{2^{*}2245.3}$

Find the solution and obtain v ₁=3411, i.e. the integrated forecasting result of the single-point load of prediction day 0: 15 (t=1) is 3411MW, and the load actual value of prediction day this point is 3376MW, and precision of prediction has reached 99.0%.

The integrated forecasting of step 6, multiple spot load:, just can obtain predicting the load sequence prediction value of day whole day for the load point repeating step two and five of other 95 moment point (t from 2 to 96) in one day day of prediction.Detailed process is identical, slightly.

Claims (3)

1, a kind of comprehensive load prediction method based on minimum information loss may further comprise the steps:

1) chooses historical reference day: take all factors into consideration various influence factors, for N historical reference day determined in a prediction day searching; If the load sampling number of every day is T, then the actual history of each historical reference day load sequence can be expressed as:

(x _n1，x _n2，…，x _nT)，n＝1，2，…，N

Wherein, x _NtThe actual load value of representing Japan-China t the sampled point of n historical reference, t=1,2 ..., T;

2) virtual prognostication: adopt M kind algorithm respectively the load of the t moment point among the load sampling number T of each historical reference day to be carried out virtual prognostication, obtain the result of virtual prognostication:

N=1,2 ..., N, m=1,2 ..., M

Represent the predicted value of m kind algorithm to Japan-China t the sampled point load of n historical reference;

3) parameter estimation:, this moment t is done the estimation of channel parameter of M kind algorithm and the estimation of information source parameter according to the result of virtual prognostication and the actual load value of historical reference day;

4) single algorithm predicts: adopt M kind algorithm respectively the t that predicts day to be loaded constantly and predict, obtain the z that predicts the outcome of each single algorithm respectively _Tm: z _TmRepresent the predicted value that m kind algorithm is loaded constantly to prediction day t;

5) integrated forecasting of single-point load: set up the objective function of information loss minimum according to step 3) and step 4) result of calculation, find the solution this objective function, obtain predicting day t v that predicts the outcome constantly _t

6) integrated forecasting of multiple spot load: for the moment point of T altogether of prediction day whole day (t value from 1 to T), repeat above step 2 respectively, just can obtain the load prediction value sequence of predicting that day whole day T is ordered: (v to step 5 ₁, v ₂..., v _T).

2, comprehensive load prediction method as claimed in claim 1 is characterized in that, the channel parameter of moment t to be estimated is the covariance matrix B of M kind algorithm channel in the described step 3) _t, the information source parameter of moment t to be estimated can comprise information source average μ _tWith the information source variances sigma _St ²

Wherein, the covariance matrix B of the M kind algorithm channel of moment t _tEstimation formulas be:

In the formula, σ _Ctm ²Expression is the variance of the m kind algorithm channel of t constantly, Cov _Ct(i, j) i of expression moment t and the covariance of j kind algorithm interchannel adopt following formula to estimate respectively:

${\mathrm{\&sigma;}}_{\mathrm{Ctm}}^2=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{({\hat{x}}_{\mathrm{ntm}}-x_{\mathrm{nt}})}^2,t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,T,m=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,M$

${\mathrm{Cov}}_{\mathrm{Ct}}(i,j)=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}({\hat{x}}_{\mathrm{nti}}-x_{\mathrm{nt}})({\hat{x}}_{\mathrm{ntj}}-x_{\mathrm{nt}}),t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,T,i,j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,M$ And i ≠ j

The information source average μ of moment t _tWith the information source variances sigma _St ²Adopt following formula to estimate respectively:

${\mathrm{\&mu;}}_t=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{x}_{\mathrm{nt}},t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,T$

${\mathrm{\&sigma;}}_{\mathrm{St}}^2=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(x_{\mathrm{nt}}{-\mathrm{\&mu;}}_t\right)}^2,t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,T.$

3, comprehensive load prediction method as claimed in claim 1 is characterized in that, the objective function of the information loss minimum in the described step 5) comprises:

When the channel Normal Distribution, information source is obeyed when evenly distributing, and the objective function of t is constantly:

$\min I_{\mathrm{loss}}(V_t;Z_t)=\frac{1}{2}{(Z_t-V_t)}^T{B}_t^{-1}(Z_t-V_t)$

In the formula, V _t=[v _t, v _t..., v _t] T is the information source vector of the M dimension of t constantly, v _tBe estimated value to the load to be predicted of moment t, Z _t=[z _T1, z _T2..., z _Ti..., z _TM] ^TBe the stay of two nights vector of the M dimension of t constantly, z _TiBe the predicting the outcome of i kind load prediction algorithm of t constantly; I _Loss(V _tZ _t) overall information loss when expression is carried out comprehensive load prediction to moment t;

When the channel Normal Distribution, during the information source Normal Distribution, the objective function of t is constantly:

$\min I_{\mathrm{loss}}(V_t;Z_t)=\frac{1}{2}{(Z_t-V_t)}^T{B}_t^{-1}(Z_t-V_t)+{(v_t-{\mathrm{\&mu;}}_t)}^2\frac{1}{2{\mathrm{\&sigma;}}_{\mathrm{St}}^2}.$

Images