CN105678426A - Method for selecting optimal day number combination of similar days in baseline prediction - Google Patents

Abstract

The invention discloses a method for selecting the optimal day number combination of similar days in baseline prediction, and the method comprises the steps: 1, carrying out the pre-screening of similar days and the selection of a similar day number combination; 2, carrying out the baseline load prediction of a similar day number combination matrix; 3, carrying out the selection of an optimal similar day number combination in the similar day number combination matrix. The method provided by the invention makes full, reasonable and effective use of a historical day number database, fully excavates non-demand response factor, quantifies the historical data similarity, carries out the ordering, and finally determines the optimal similar day number combination through the error comparison of different day number combinations. The method employs the advantages of a gray correlation method, a comprehensive weighting method, an MLR (Multiple Linear Regression) method, and an MRE error evaluation method, greatly improves the economic performance and accuracy of demand response baseline prediction, and provides a scientific theoretical support for the technology of demand response.

Description

The choosing method of the optimum number of days combination of similar day in a kind of baseline forecast

Technical field

The present invention relates to the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast, belong to Power System and its Automation technical field.

Background technology

Since new millennium; electricity needs rapidly increases; although power construction high speed development; but the growth at full speed due to workload demand; energy dilemma and imbalance between supply and demand are increasingly serious, short of electricity and the peak-valley difference enlargement phenomenon of China's most area region property, seasonality, period, structure and by long-term existence. Demand response is the important technical of dsm, refers to that price or actuation signal are made response by user, and changes normal electricity consumption mode, thus realizes configuring with the electrically optimized complex optimum with system resource. But in the practice that we are concrete, baseline forecast for demand response is the big difficult problem existed always, common baseline forecast method is chosen similar day based on factors such as user's day type, temperature and is predicted, but often specifically do not quantize the similarity of history day, the concrete number of days of similar day is not determined yet, causing cannot rationally and efficiently utilizing historical data, accuracy and economy to baseline forecast cannot ensure. The multiple factors how obtained by data mining rationally carry out the sequence of history day similarity, and determine specifically the choosing in the calculating that number of days brings baseline load into of similar day, the place do not considered when being demand response baseline forecast in the past.

Summary of the invention

Object: in order to overcome the deficiencies in the prior art, the present invention provides the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast.

Technical scheme: for solving the problems of the technologies described above, the technical solution used in the present invention is:

In baseline forecast, a choosing method for the optimum number of days combination of similar day, comprises the steps:

Step one: the pre-screening of similar day and the combination of similar day number of days are chosen;

1a: determine that choosing scope history day is 28 days, by the load curve of a period before the demand response period and demand response day with period load curve dependency, compute associations degree γ_i, select γ_iThe history day of > 0.9 is similar day;

1b: in size of a sample s+28, by step 1a requirement, chooses similar day sequence { d₁,d₂,……,d_s, the minimum value got in sequence is designated as d, chooses number of days as similar day combination is maximum;

1c: to assume s demand response day each correspondence d similar day carry out the sequence of comprehensive similarity, and choose different similar day number of days combinations according to sequence, what each was assumed obtains d kind scheme demand response day, namely forms the similar day combinatorial matrix of a s × d $A=\left[\begin{array}{cccc}{a}_1\left(1\right)& a_2\left(1\right)& \mathrm{...}& a_d\left(1\right)\\ a_1\left(2\right)& a_2\left(2\right)& \mathrm{...}& a_d\left(2\right)\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ a_1\left(s\right)& a_2\left(s\right)& \mathrm{...}& a_d\left(s\right)\end{array}\right];$

Step 2: be assumed to be for s+28 days with and adopt d kind scheme demand response day, i.e. a in similar day combinatorial matrix A_dS () carries out the prediction of baseline load for example;

2a: the weighting coefficient calculating similar dayF_mS () is the comprehensive similarity using the s+28 days as m similar day during demand response day

2b: calculate correction factor $c_m\left(s\right)=\frac{\underset{j=I-T}{\overset{I-1}{\Σ}}{P}_{s+28}(s,j)}{\underset{j=I-T}{\overset{I-1}{\Σ}}{P}_m(s,j)},m=1,2,\mathrm{...},d;$

2c: calculate the similar day load P' after revising_m(s, j)=c_m(s)×P_m(s, j) m=1,2 ..., d; J=I, I+1 ..., I+J;

2d: calculate baseline load predictor $P_{basic,d}(s,j)=\underset{m=1}{\overset{d}{\Σ}}{\mathrm{\ω}}_m\left(s\right){P^{\′}}_m(s,j)$ M=1,2 ..., d; J=I, I+1 ..., I+J;

2e: the baseline load predictor matrix obtaining d kind scheme

$\[P_{ba\mathrm{si}c,d}\]=\left[\begin{array}{cccc}{P}_{basic,d}(1,I)& P_{basic,d}(1,I+1)& \mathrm{...}& P_{basic,d}(1,I+J)\\ P_{basic,d}(2,I)& P_{basic,d}(2,I+1)& \mathrm{...}& P_{basic,d}(2,I+J)\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ P_{basic,d}(s,I)& P_{basic,d}(s,I+1)& \mathrm{...}& P_{basic,d}(s,I+J)\end{array}\right];$

Step 3: in similar day combinatorial matrix A, optimum similar day number of days combination is chosen;

3a: calculate often kind of scheme average relative error in similar day combinatorial matrix A

${\mathrm{MRE}}_l=\frac{{\mathrm{MRE}}_l\left(1\right)+{\mathrm{MRE}}_l\left(2\right)+\mathrm{...}+{\mathrm{MRE}}_l\left(s\right)}{s};$

The MRE of 3b: more often kind scheme_l, choose MRE_lSky numerical value in minimum scheme is as optimum similar day number of days combination.

Described calculation of relationship degree method is as follows:

Step one: the time period of known enforcement demand response is I～I+J, now chooses the load curve of time period I-T～I-1 before the corresponding demand response period of each history day and demand response day, adopts grey correlation method to carry out similarity comparison;

Step 2: determine 28 history days, choose the load value that the time period is I-T～I-1 and form comparative sequences:

x_i={ x_i(1),x_i(2),x_i(3),……,x_i(T) } i=1,2 ..., 28 (1)

In formula (1), x_i(T) it is the I-1 load characteristic amount of i-th history day;

Step 3: the load characteristic amount sequence obtaining the demand response day corresponding period, is designated as reference sequences y:

Y={y (1), y (2), y (3) ..., y (T) } and (2)

In formula (2), the I-1 load characteristic amount that y (T) is demand response day;

Step 4: compute associations coefficient:

${\mathrm{\ξ}}_i\left(t\right)=\frac{\underset{1\≤i\≤28}{\min }\underset{1\≤t\≤T}{\min }|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}{|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}---\left(3\right)$

In formula (3), ξ_iT () is reference sequences y and comparative sequences x_iIncidence coefficient in the I-1 load characteristic amount, t=1,2 ..., T; ρ, for differentiating coefficient, gets ρ=0.5 usually;

Step 5: calculate y and comparative sequences x_iRelational degree be:

${\mathrm{\γ}}_i=\frac{1}{T}\underset{t=1}{\overset{T}{\Σ}}{\mathrm{\ξ}}_i\left(t\right)---\left(4\right)$

In formula (4), γ_iFor y and comparative sequences x_iRelational degree, its value characterizes the dependency of i-th history day at the load curve load in some time curve corresponding to demand response day implementing time period I-T～I-1 before demand response.

Described similar day combinatorial matrix A method of calculation are as follows:

Step one: the history day of the 29th day is assumed to be demand response day and the pre-screening that its history day of first 28 days carries out similar day is drawn d₁Individual similar day, draws d with reason the 30th day be assumed to be history day demand response day₂Individual similar day, analogizes with this, within s+28 days, obtains d to_sIndividual similar day, obtains sequence:

{d₁,d₂,……,d_s}(5)

In formula (5), d_sThe s+28 days as demand response day time, screen the similar day quantity obtained in advance;

Get the minimum value in sequence as and be designated as d, that is:

D=min{d₁,d₂,……,d_s}

Represent for from the 29th day to the s+28 days altogether s days respectively as demand response day time, at least can obtain d similar day by screening in advance, namely d is that the similar day of this sample combines and maximum chooses number of days;

Step 2: known similar day number be d, consider the non-demand response factor of N kind, the time period implementing demand response is I～I+J, calculate in I～I+J period similar day and the similarity of demand response day respectively for each factor, wherein the maximum similar value of each factor is 1, obtains similarity matrix as follows:

M=[m_nk] n=1,2 ..., N; K=1,2 ..., d (6)

In formula (6), M is similarity matrix; m_nkFor the n-th factor of kth similar day is in the similarity of I～I+J period factor corresponding to demand response day in the corresponding period;

Step 3: calculate each similar day and demand response day in the similarity of I～I+J period by M, similar day is multiplied with the weighting similarity of each factor of period, is:

$F_k=\underset{n=1}{\overset{N}{\Π}}{\[m_{nk}\]}^{R_n}---\left(7\right)$

In formula (7), F_kFor kth similar day is in the comprehensive similarity of I～I+J period; R_nIt is the Intrusion Index of n-th kind of factor, R_nMore big this factor of expression is more big to the influence degree of load;

Step 4: calculate the comprehensive similarity of d similar day respectively and the comprehensive similarity obtained is carried out sequence from big to small;

Step 5: the demand response day assumed for one, the history day choosing sequencing of similarity the 1st is designated as scheme 1 as similar day number of days assembled scheme, is designated as a₁; The history day choosing first 2 of sequencing of similarity is designated as scheme 2 as number of days assembled scheme, is designated as a₂; Analogizing with this, obtaining number of days assembled scheme sequence is:

{a₁,a₂,……,a_d}(8)

In formula (8), a_dBeing the d scheme, the history day namely choosing before sequencing of similarity d is as similar day;

To s the demand response day assumed, there is d kind scheme respectively, namely form the similar day number of days combinatorial matrix of a s × d:

$A=\left[\begin{array}{cccc}{a}_1\left(1\right)& a_2\left(1\right)& ...& a_d\left(1\right)\\ a_1\left(2\right)& a_2\left(2\right)& ...& a_d\left(2\right)\\ ...& ...& ...& ...\\ a_1\left(s\right)& a_2\left(s\right)& ...& a_d\left(s\right)\end{array}\right]---\left(9\right)$

In formula (9), a_dS () characterizes the characteristic quantity using the s+28 days as d scheme during demand response day.

Described average relative error method of calculation are as follows:

Step one: the baseline load predictor scheme matrix [P of the d kind scheme that can obtain respectively by above-mentioned steps_basic,l], l=1,2 ..., d, the actual negative charge values contrasting known hypothesis demand response day calculates;

Step 2: the calculation formula of average relative error MRE index is:

${\mathrm{MRE}}_l\left(s\right)=\frac{1}{J+1}\underset{j=I}{\overset{I+J}{\Σ}}\left|\frac{P_{s+28}(s,j)-P_{basic,l}(s,j)}{P_{s+28}(s,j)}\right|\×100\%,l=1,2,......,d---\left(10\right)$

In formula (10), MRE_lThe average relative error of l kind scheme when () is for assuming by demand response day of s+28 s; P_s+28(s, j) is the load value using the s+28 days as the jth time period on same day during demand response day; P_basic,l(s, j) is the baseline load predictor of the j period obtained as l kind scheme during demand response day using the s+28 days;

Step 3: the average relative error value of s sample is averaged and obtains often kind of scheme baseline forecast relative error magnitudes and be:

${\mathrm{MRE}}_l=\frac{{\mathrm{MRE}}_l\left(1\right)+{\mathrm{MRE}}_l\left(2\right)+...+{\mathrm{MRE}}_l\left(s\right)}{s}---\left(11\right)$

In formula (11), MRE_lIt it is the baseline load Relative Error value of l kind scheme.

Useful effect: the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast provided by the invention, fully, rationally and efficiently utilize history day database, fully excavate non-demand response factor, quantification history day similarity also sorts, the error contrast combined by different number of days is finally determined similar day optimum and chooses number of days combination, wherein have employed grey correlation method, aggregative weighted method, multiple linear regression method, the advantage of the methods such as MRE error assessment, drastically increase economy and the accuracy of demand response baseline forecast, for demand response technology provides scientific theory support.

Accompanying drawing explanation

Fig. 1 is the general flow chart of the inventive method.

Embodiment

Below in conjunction with accompanying drawing, the present invention is further described.

As shown in Figure 1, the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast:

One, the pre-screening of similar day

The influence factor that 1.1 similar day are selected

In order to similar day being carried out reasonable effective pre-screening, first wanting its influence factor clear and definite, now analyzing its influence factor from following 5 aspects.

(1) range of choice of history day

Generally, the history day similarity more near apart from day to be measured is more high, also passing in time slowly change can be there is in system loading structure, therefore select influence factor very similar but the timed interval is longer history day its prediction effect can not be desirable, excessive range of choice also can increase the cost that historical data is excavated. Meanwhile, when selecting similar day, having significantly " near big and far smaller " rule, the similarity gap of be namely separated by with day to be measured 1 day and be separated by history day of 1 week and day to be measured is bigger; But with history day its similarity not obviously difference be separated by day to be measured 3 weeks and be separated by 4 weeks. Exact date, the similarity formula of distance was as follows:

$m_{ik}=\left\{\begin{array}{c}{\mathrm{\β}}^k,{\mathrm{\β}}^k\≥a\\ a\end{array}\right.---\left(1\right)$

In formula (1), m_ikFor the similarity of kth sky date distance factor; β is reduction coefficient, usually gets between 0.9～0.98; A is the minimum similarity of date distance factor, usually gets between 0.2～0.4.

Therefore, herein when selecting similar day, by front surrounding, the history day of 28 days is as similar day scope to be selected.

(2) day type factor

Day type refers to single prediction day feature just can known from the load date. Short-term electric load has obvious date periodicity and weekly pattern; Generally, weekend, load was significantly lower than working days, and load curve shape also has obvious difference. In working days, load curve shape is substantially identical, but Monday, load can be different from general work daily load by the impact of all daily loads, and Friday, load also can be different from general work daily load because of the arrival at weekend. When day, type was identical, obtaining this factor maximum similarity is 1; Be all working days but day the different situation of type, provide a similarity, such as 0.7; Similarity between working days and non-working days gets 0; Etc..

For the special day such as the Spring Festival, New Year's Day, predicting as similar day history day by same special day type, does not do to consider herein usually.

(3) meteorological factor

Meteorological factor comprises the temperature of load day, the highest, minimum or real-time, and humidity, rainfall amount, sunshine, air pressure, weather pattern etc. are the one of the main reasons affecting daily load curve. To be ensured availability and the data quantification property of its historical data when obtaining meteorological factor, be convenient to the judgement criteria as comprehensive similarity.

(4) load average

Daily load average can reflect the load level on the same day to a certain extent, is the important factor during baseline load is predicted. But we are when baseline forecast, normally load for a certain the period of load Japan and China is predicted, instead of the prediction of whole day load total amount, therefore to be carried out analyses and prediction in conjunction with its load curve.

(5) similarity of load curve

Article two, similar " the vertical translation similarity " referring to curve of the shape of load curve, namely the greatest coincide degree of two curves when vertical direction translates, but it is noted that only the overall shape similarity height of load curve can not represent its local shape similarity height, we are when baseline forecast, usually to be analyzed for the local load curve similarity of a certain period.

The pre-screening of 1.2 similar day and the combination of similar day number of days are chosen

The history day determined by front literary composition chooses scope 28 days, is first carried out the pre-screening of similar day by the load curve dependency of a period before the demand response period, then the similar day comprehensive similarity evaluation function according to each non-demand response factors composition sorts.

(1) the pre-screening of similar day

User's baseline load refers to if load when user does not participate in demand response project, we are when to this section of load curve forecasting, need with reference to implementing the load curve before the demand response period, similarity is contrasted from history daily load curve, carry out similar day to screen in advance, thus infer baseline load curve.

Being now the history day of s+28 for size of a sample, within s+28 days, be assumed to be demand response day to the 29th day to the respectively, before corresponding demand response day 28 days was then history day, carried out the pre-screening of similar day. If the 29th day is assumed to be demand response day, it it within the 1-28 days, it is then history day; If the 30th day is assumed to be demand response day, within the 2-29 days, it is then history day, analogizes with this.

The time period of known enforcement demand response is I～I+J, has 96 instance sample points every day, and the load curve now choosing time period I-T～I-1 before the corresponding demand response period of each history day and demand response day adopts grey correlation method to carry out similarity comparison.

To 28 history days, choose the load value that the time period is I-T～I-1 and form comparative sequences:

x_i={ x_i(1),x_i(2),x_i(3),……,x_i(T) } i=1,2 ..., 28 (2)

In formula (2), x_i(T) it is the I-1 load characteristic amount of i-th history day;

Obtain the load characteristic amount sequence of demand response day corresponding period, it be designated as reference sequences y:

Y={y (1), y (2), y (3) ..., y (T) } and (3)

In formula (3), the I-1 load characteristic amount that y (T) is demand response day.

Compute associations coefficient:

${\mathrm{\ξ}}_i\left(t\right)=\frac{\underset{1\≤i\≤28}{\min }\underset{1\≤t\≤T}{\min }|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}{|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}---\left(4\right)$

In formula (4), ξ_iT () is reference sequences y and comparative sequences x_iIncidence coefficient in the I-1 load characteristic amount, t=1,2 ..., T; ρ, for differentiating coefficient, gets ρ=0.5 usually.

Calculate y and comparative sequences x_iRelational degree be:

${\mathrm{\γ}}_i=\frac{1}{T}\underset{t=1}{\overset{T}{\Σ}}{\mathrm{\ξ}}_i\left(t\right)---\left(5\right)$

In formula, γ_iFor y and comparative sequences x_iRelational degree, its value characterizes the dependency of i-th history day at the load curve load in some time curve corresponding to demand response day implementing time period I-T～I-1 before demand response.

Herein, select the relational degree γ with y_iThe history day of load curve corresponding to the comparative sequences of > 0.9 is as similar day.

(2) combination of sample similar day number of days is chosen

The history day of the 29th day is assumed to be demand response day and the pre-screening that its history day of first 28 days carries out similar day is drawn d₁Individual similar day, draws d with reason the 30th day be assumed to be history day demand response day₂Individual similar day, analogizes with this, within s+28 days, obtains d to_sIndividual similar day, obtains sequence:

{d₁,d₂,……,d_s}(6)

In formula (6), d_sThe s+28 days as demand response day time, screen the similar day quantity obtained in advance.

Get the minimum value in sequence as and be designated as d, that is:

D=min{d₁,d₂,……,d_s}(7)

For from the 29th day to the s+28 days altogether s days respectively as demand response day time, at least can obtain d similar day by screening in advance, namely d is that the similar day of this sample combines and maximum chooses number of days.

Two, the baseline forecast of different similar day number of days combination

The above-mentioned stage determines d and demand response daily load relational degree bigger history day altogether as similar day, and d similar day will be carried out the sequence of comprehensive similarity and choose different similar day number of days and combine by this stage, carry out the prediction of baseline load.

The calculating of 2.1 comprehensive similarity and sequence

Existing known is d at similar day number, this algorithm is considered N kind non-demand response factor altogether, the time period implementing demand response is I～I+J, calculate in I～I+J period similar day and the similarity of day to be measured respectively for each factor, wherein the maximum similar value of each factor is 1, obtains similarity matrix as follows:

M=[m_nk] n=1,2 ..., N;K=1,2 ..., d (8)

In formula (8), M is similarity matrix; m_nkFor the n-th factor of kth similar day is in the similarity of I～I+J period factor corresponding to demand response day in the corresponding period.

And calculate each similar day and day to be measured in the similarity of I～I+J period by M, similar day is multiplied with the weighting similarity of each factor of period, is:

$F_k=\underset{n=1}{\overset{N}{\Π}}{\[m_{nk}\]}^{R_n}---\left(9\right)$

In formula (9), F_kFor kth history day is in the comprehensive similarity of I～I+J period; R_nIt is the Intrusion Index of n-th kind of factor, R_nMore big this factor of expression is more big to the influence degree of load.

Calculate the comprehensive similarity of d similar day respectively and the comprehensive similarity obtained is carried out sequence from big to small.

2.2 structure similar day combinatorial matrixs

For the demand response day that one is assumed, the history day choosing sequencing of similarity the 1st is designated as scheme 1 as similar day number of days assembled scheme, is designated as a₁; The history day choosing first 2 of sequencing of similarity is designated as scheme 2 as number of days assembled scheme, is designated as a₂; Analogizing with this, obtaining number of days assembled scheme sequence is:

{a₁,a₂,……,a_d}(10)

In formula (10), a_dBeing the d scheme, the history day namely choosing before sequencing of similarity d is as similar day.

To s the demand response day assumed, there is d kind scheme respectively, namely form the similar day number of days combinatorial matrix of a s × d:

$A=\left[\begin{array}{cccc}{a}_1\left(1\right)& a_2\left(1\right)& \mathrm{...}& a_d\left(1\right)\\ a_1\left(2\right)& a_2\left(2\right)& \mathrm{...}& a_d\left(2\right)\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ a_1\left(s\right)& a_2\left(s\right)& \mathrm{...}& a_d\left(s\right)\end{array}\right]---\left(11\right)$

In formula (11), a_dS () characterizes the characteristic quantity using the s+28 days as d scheme during demand response day.

2.3 baseline load predictions

Similar day weighted method

Adopt d kind scheme using the s+28 days as demand response day now, i.e. a in similar day combinatorial matrix A_dS () carries out the prediction of baseline load for example. The comprehensive similarity sequence of d similar day being obtained choosing by Similarity Measure previous stage is { F₁(s),F₂(s),……,F_d(s) }, the weighting coefficient that thus can obtain similar day is:

${\mathrm{\ω}}_m\left(s\right)=\frac{F_m\left(s\right)}{\underset{i=1}{\overset{d}{\Σ}}F\left(s\right)},m=1,2,......,d---\left(12\right)$

In formula (12), F_mS () is the comprehensive similarity using the s+28 days as m similar day during demand response day, ω_mS () is the weighting coefficient using the s+28 days as m similar day during demand response day.

Owing to only considered the degree of correlation of load curve in the process screened in advance at similar day, not considering the impact of load average, to be carried out considering the corrected Calculation of load average here, the expression formula of its correction factor is:

$c_m\left(s\right)=\frac{\underset{j=I-T}{\overset{I-1}{\Σ}}{P}_{s+28}(s,j)}{\underset{j=I-T}{\overset{I-1}{\Σ}}{P}_m(s,j)},m=1,2,...,d---\left(13\right)$

In formula (13), c_mS () is the correction factor using the s+28 days as the m similar day load during demand response day; P_mWhen (s, j) is using the s+28 days as demand response day, the m similar day is at the load value of jth time period; P_s+28When (s, j) is using the s+28 days as demand response day, the load value of the jth time period on the same day.

Therefore the similar day load after can being revised is:

P'_m(s, j)=c_m(s)×P_m(s, j) m=1,2 ..., d; J=I, I+1 ..., I+J (14)

In formula (14), P'_m(s, j) is load value after the correction of jth time period using kth+28 days as the m similar day during demand response day.

Obtaining baseline load predictor according to the load value after correction and its weighting coefficient is:

$P_{basic,d}(s,j)=\underset{m=1}{\overset{d}{\Σ}}{\mathrm{\ω}}_m\left(s\right){P^{\′}}_m(s,j),m=1,2,\mathrm{...},d;j=I,I+1,\mathrm{...},I+J---\left(15\right)$

In formula, P_basic,d(s, j) is the baseline load predictor of the j period obtained as d kind scheme during demand response day using the s+28 days.

The baseline load predictor matrix obtaining d kind scheme is:

$\[P_{ba\mathrm{si}c,d}\]=\left[\begin{array}{cccc}{P}_{basic,d}(1,I)& P_{basic,d}(1,I+1)& \mathrm{...}& P_{basic,d}(1,I+J)\\ P_{basic,d}(2,I)& P_{basic,d}(2,I+1)& \mathrm{...}& P_{basic,d}(2,I+J)\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ P_{basic,d}(s,I)& P_{basic,d}(s,I+1)& \mathrm{...}& P_{basic,d}(s,I+J)\end{array}\right]---\left(16\right)$

3 error contrasts and number of days assembled scheme are determined

Scheme matrix [the P of the d kind scheme that can obtain respectively by above-mentioned steps_basic,l], l=1,2 ..., d, contrasts the actual negative charge values of known hypothesis demand response day, adopts average relative error MRE index to judge as the judging basis of prediction precision, obtains optimum similar day number of days combination.

3.1 average relative error MRE

The calculation formula of this index is:

${\mathrm{MRE}}_l=\frac{1}{J+1}\underset{j=I}{\overset{I+J}{\Σ}}\left|\frac{P_{s+28}(s,j)-P_{basic,l}(s,j)}{P_{s+28}(s,j)}\right|\×100\%,l=1,2,\mathrm{......},d---\left(17\right)$

In formula (17), MRE_lThe average relative error of l kind scheme when () is for assuming by demand response day of s+28 s;P_s+28(s, j) is the load value using the s+28 days as the jth time period on same day during demand response day; P_basic,l(s, j) is the baseline load predictor of the j period obtained as l kind scheme during demand response day using the s+28 days.

The average relative error value of s sample is averaged and obtains often kind of scheme baseline forecast relative error magnitudes and be:

${\mathrm{MRE}}_l=\frac{{\mathrm{MRE}}_l\left(1\right)+{\mathrm{MRE}}_l\left(2\right)+...+{\mathrm{MRE}}_l\left(s\right)}{s}---\left(18\right)$

In formula (18), MRE_lIt it is the baseline load Relative Error value of l kind scheme.

The determination of 3.2 optimum number of days combinations

MRE is compared respectively for d kind scheme_l, choose and make MRE_lMinimum l chooses number of days as optimum similar day.

The above is only the preferred embodiment of the present invention; it is noted that, for those skilled in the art; under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (4)

1. the choosing method of the optimum number of days combination of similar day in a baseline forecast, it is characterised in that: comprise the steps:

Step one: the pre-screening of similar day and the combination of similar day number of days are chosen;

1a: determine that choosing scope history day is 28 days, by the load curve of a period before the demand response period and demand response day with period load curve dependency, compute associations degree γ_i, select γ_iThe history day of > 0.9 is similar day;

1b: in size of a sample s+28, by step 1a requirement, chooses similar day sequence { d₁,d₂,……,d_s, the minimum value got in sequence is designated as d, chooses number of days as similar day combination is maximum;

1c: to assume s demand response day each correspondence d similar day carry out the sequence of comprehensive similarity, and choose different similar day number of days combinations according to sequence, what each was assumed obtains d kind scheme demand response day, namely forms the similar day combinatorial matrix of a s × d $A=\left[\begin{array}{cccc}{a}_1\left(1\right)& a_2\left(1\right)& ...& a_d\left(1\right)\\ a_1\left(2\right)& a_2\left(2\right)& ...& a_d\left(2\right)\\ ...& ...& ...& ...\\ a_1\left(s\right)& a_2\left(s\right)& ...& a_d\left(s\right)\end{array}\right];$

Step 2: be assumed to be for s+28 days with and adopt d kind scheme demand response day, i.e. a in similar day combinatorial matrix A_dS () carries out the prediction of baseline load for example;

2a: the weighting coefficient calculating similar dayM=1,2 ..., d; F_mS () is the comprehensive similarity using the s+28 days as m similar day during demand response day

2b: calculate correction factorM=1,2 ..., d;

2c: calculate the similar day load P' after revising_m(s, j)=c_m(s)×P_m(s, j) m=1,2 ..., d; J=I, I+1 ..., I+J;

2d: calculate baseline load predictor $P_{basic,d}(s,j)=\underset{m=1}{\overset{d}{\Σ}}{\mathrm{\ω}}_m\left(s\right){P^{\′}}_m(s,j)$ M=1,2 ..., d; J=I, I+1 ..., I+J;

2e: the baseline load predictor matrix obtaining d kind scheme

$\[P_{basic,d}\]=\left[\begin{array}{cccc}{P}_{basic,d}(1,I)& P_{basic,d}(1,I+1)& ...& P_{basic,d}(1,I+J)\\ P_{basic,d}(2,I)& P_{basic,d}(2,I+1)& ...& P_{basic,d}(2,I+J)\\ ...& ...& ...& ...\\ P_{basic,d}(s,I)& P_{basic,d}(s,I+1)& ...& P_{basic,d}(s,I+J)\end{array}\right];$

Step 3: in similar day combinatorial matrix A, optimum similar day number of days combination is chosen;

3a: calculate often kind of scheme average relative error in similar day combinatorial matrix A

${\mathrm{MRE}}_l=\frac{{\mathrm{MRE}}_l\left(1\right)+{\mathrm{MRE}}_l\left(2\right)+...+{\mathrm{MRE}}_l\left(s\right)}{s};$

The MRE of 3b: more often kind scheme_l, choose MRE_lSky numerical value in minimum scheme is as optimum similar day number of days combination.

2. the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast according to claim 1, it is characterised in that: described calculation of relationship degree method is as follows:

Step one: the time period of known enforcement demand response is I～I+J, now chooses the load curve of time period I-T～I-1 before the corresponding demand response period of each history day and demand response day, adopts grey correlation method to carry out similarity comparison;

Step 2: determine 28 history days, choose the load value that the time period is I-T～I-1 and form comparative sequences:

x_i={ x_i(1),x_i(2),x_i(3),……,x_i(T) } i=1,2 ..., 28 (1)

In formula (1), x_i(T) it is the I-1 load characteristic amount of i-th history day;

Step 3: the load characteristic amount sequence obtaining the demand response day corresponding period, is designated as reference sequences y:

Y={y (1), y (2), y (3) ..., y (T) } and (2)

In formula (2), the I-1 load characteristic amount that y (T) is demand response day;

Step 4: compute associations coefficient:

${\mathrm{\ξ}}_i\left(t\right)=\frac{\underset{1\≤i\≤28}{min}\underset{1\≤t\≤T}{min}|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}{|y\left(t\right)-x_i\left(t\right)|+\mathrm{\ρ}\underset{1\≤i\≤28}{\max }\underset{1\≤t\≤T}{\max }|y\left(t\right)-x_i\left(t\right)|}---\left(3\right)$

In formula (3), ξ_iT () is reference sequences y and comparative sequences x_iIncidence coefficient in the I-1 load characteristic amount, t=1,2 ..., T; ρ, for differentiating coefficient, gets ρ=0.5 usually;

Step 5: calculate y and comparative sequences x_iRelational degree be:

${\mathrm{\γ}}_i=\frac{1}{T}\underset{t=1}{\overset{T}{\Σ}}{\mathrm{\ξ}}_i\left(t\right)---\left(4\right)$

In formula (4), γ_iFor y and comparative sequences x_iRelational degree, its value characterizes the dependency of i-th history day at the load curve load in some time curve corresponding to demand response day implementing time period I-T～I-1 before demand response.

3. the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast according to claim 1, it is characterised in that: described similar day combinatorial matrix A method of calculation are as follows:

Step one: the history day of the 29th day is assumed to be demand response day and the pre-screening that its history day of first 28 days carries out similar day is drawn d₁Individual similar day, draws d with reason the 30th day be assumed to be history day demand response day₂Individual similar day, analogizes with this, within s+28 days, obtains d to_sIndividual similar day, obtains sequence:

{d₁,d₂,……,d_s}(5)

In formula (5), d_sThe s+28 days as demand response day time, screen the similar day quantity obtained in advance;

Get the minimum value in sequence as and be designated as d, that is:

D=min{d₁,d₂,……,d_s}

Represent for from the 29th day to the s+28 days altogether s days respectively as demand response day time, at least can obtain d similar day by screening in advance, namely d is that the similar day of this sample combines and maximum chooses number of days;

Step 2: known similar day number be d, consider the non-demand response factor of N kind, the time period implementing demand response is I～I+J, calculate in I～I+J period similar day and the similarity of demand response day respectively for each factor, wherein the maximum similar value of each factor is 1, obtains similarity matrix as follows:

M=[m_nk] n=1,2 ..., N; K=1,2 ..., d (6)

In formula (6), M is similarity matrix; m_nkFor the n-th factor of kth similar day is in the similarity of I～I+J period factor corresponding to demand response day in the corresponding period;

Step 3: calculate each similar day and demand response day in the similarity of I～I+J period by M, similar day is multiplied with the weighting similarity of each factor of period, is:

$F_k=\underset{n=1}{\overset{N}{\Π}}{\[m_{nk}\]}^{R_n}---\left(7\right)$

In formula (7), F_kFor kth similar day is in the comprehensive similarity of I～I+J period; R_nIt is the Intrusion Index of n-th kind of factor, R_nMore big this factor of expression is more big to the influence degree of load;

Step 4: calculate the comprehensive similarity of d similar day respectively and the comprehensive similarity obtained is carried out sequence from big to small;

Step 5: the demand response day assumed for one, the history day choosing sequencing of similarity the 1st is designated as scheme 1 as similar day number of days assembled scheme, is designated as a₁; The history day choosing first 2 of sequencing of similarity is designated as scheme 2 as number of days assembled scheme, is designated as a₂; Analogizing with this, obtaining number of days assembled scheme sequence is:

{a₁,a₂,……,a_d}(8)

In formula (8), a_dBeing the d scheme, the history day namely choosing before sequencing of similarity d is as similar day; To s the demand response day assumed, there is d kind scheme respectively, namely form the similar day number of days combinatorial matrix of a s × d:

$A=\left[\begin{array}{cccc}{a}_1\left(1\right)& a_2\left(1\right)& ...& a_d\left(1\right)\\ a_1\left(2\right)& a_2\left(2\right)& ...& a_d\left(2\right)\\ ...& ...& ...& ...\\ a_1\left(s\right)& a_2\left(s\right)& ...& a_d\left(s\right)\end{array}\right]---\left(9\right)$

In formula (9), a_dS () characterizes the characteristic quantity using the s+28 days as d scheme during demand response day.

4. the choosing method of the optimum number of days combination of similar day in a kind of baseline forecast according to claim 1, it is characterised in that: described average relative error method of calculation are as follows:

Step one: the baseline load predictor scheme matrix [P of the d kind scheme that can obtain respectively by above-mentioned steps_basic,l], l=1,2 ..., d, the actual negative charge values contrasting known hypothesis demand response day calculates;

Step 2: the calculation formula of average relative error MRE index is:

$\begin{array}{cc}{\mathrm{MRE}}_l\left(s\right)=\frac{1}{J+1}\underset{j=I}{\overset{I+J}{\mathrm{\Σ}}}\left|\frac{P_{s+28}(s,j)-P_{basic,l}(s,j)}{P_{s+28}(s,j)}\right|\×100\%& l=1,2,......,d\end{array}---\left(10\right)$

In formula (10), MRE_lThe average relative error of l kind scheme when () is for assuming by demand response day of s+28 s; P_s+28(s, j) is the load value using the s+28 days as the jth time period on same day during demand response day; P_basic,l(s, j) is the baseline load predictor of the j period obtained as l kind scheme during demand response day using the s+28 days;

Step 3: the average relative error value of s sample is averaged and obtains often kind of scheme baseline forecast relative error magnitudes and be:

${\mathrm{MRE}}_l=\frac{{\mathrm{MRE}}_l\left(1\right)+{\mathrm{MRE}}_l\left(2\right)+...+{\mathrm{MRE}}_l\left(s\right)}{s}---\left(11\right)$

In formula (11), MRE_lIt it is the baseline load Relative Error value of l kind scheme.