CN105184388A - Non-linear regression method for urban power load short-period prediction - Google Patents

Abstract

The invention relates to a non-linear regression method for urban power load short-period prediction. A relation curve influencing a factor change is drawn; according to industrial, commercial and resident load proportionality coefficients based on statistics, poly-type non-linear composite model characteristic curves of all loads are obtained by analyses; an accurate multi-element poly-type composite non-linear regression model is obtained; and for the obtained multi-element poly-type composite non-linear regression model, a model prediction value is calculated based on hypothesis testing in mathematical statistics by using a measured load data sample, verification is carried out according to characteristics of all loads obtained by analyses in advance, and determination whether to refuse or accept the model hypothesis is made, thereby obtaining quantitative data related to the model credibility. According to the invention, all factors influencing the load change can be added into the prediction model by using a multi-element regression model; and with the non-linear regression model, defects of the linear regression model can be effectively overcome, thereby improving the prediction precision.

Description

A kind of non-linear regression method of urban power load short-term forecasting

Technical field

The non-linear regression method of a kind of urban power load short-term forecasting of the present invention, relates to urban power load short-term forecasting field.

Background technology

Urban power load short-term forecasting refers generally to the load index of following a day of predicted city or a week, and its meaning is to help yardman to determine fuel supply plan; Operating power plant is exerted oneself and requires to propose advance notice.Make to be estimated in advance the state change of generator; The start and stop of each unit in Home Network can be arranged economically, reduce and rotate idle capacity; Can when ensureing normal electricity consumption reasonable arrangement unit maintenance scheduling.Urban power load is along with many factors generation nonlinearities change, and first load forecasting method will meet this polynary and nonlinear relevance; Secondly, load forecasting method requires to have relative stability in practicality, and its precision of prediction with the change of predicted time, big ups and downs does not occur.

Urban electric power short-term load forecasting method has experience forecasting techniques, linear regression prediction technology, time series forecasting technology and intelligent predicting technology four class.

1), rule-of-thumb prediction technique mainly relies on the judgement of expert or expert group, not rely on quantitative model, object is not the track and structure of understanding fully that electric load changes, but provide the conclusion of a directivity, the defect of the method is that predicated error is higher, therefore actually seldom to adopt, only needs according to a preliminary estimate load variations time just employing.

2), linear regression prediction refers to, with the regression analysis in mathematical statistics, namely by carrying out statistical study to the observation data of variable, determine the linear relationship between variable, thus realizes the object of prediction.The defect of the method is as follows: (1), nonlinear relationship often between actual electric load and variation, and linear regression model (LRM) cannot react this relation; (2), the method cannot carry out classified synthesis prediction to electric load, is thus not suitable for different city load predictions.

3), time series forecasting technology is different from linear regression technique, and dependent variable (target of prediction) and independent variable can be all stochastic variables.As the electric load that dependent variable is now to be measured, and independent variable is the past value of load self.Now, dependent variable and independent variable are all stochastic variables.The major defect of the method is cannot with the variable association affecting load variations.

4), intelligent predicting technology mainly refers to Artificial Neural Network, and the nerve network system of simulation human brain, obtains suitable parameter by study, is used for the nonlinear relationship that mapping pair answers.The method has self study, self-adaptation and Integrated Fault Tolerant ability, but defect is that prediction is unstable, indigestion and being theoretically unsound.

Summary of the invention

For solving the problems of the technologies described above, the invention provides a kind of non-linear regression method of urban power load short-term forecasting, the factor affecting urban power load is divided into the meteorological conditions such as the temperature of every day, wind speed and weather, and several class such as date feature, affect between a few class variable of load variations and load in nonlinear relationship; Meanwhile, urban power load generally comprises three basic forms of it: industrial load, Commercial Load and appliance load.These dissimilar loads are set up nonlinear relationship model between different load and each variation according to historical load data, then polytype load is compounded to form the multivariate nonlinear regression analysis model of city load entirety, adopts method for parameter estimation to ask for each design parameter of model.The method goes for the city short-term electric load prediction of China different regions, different development level, and precision of prediction is high, prediction is stable.

The technical solution adopted in the present invention is:

A non-linear regression method for urban power load short-term forecasting, comprises the following steps:

Step one: carry out load Analysis: the load of quantitative statistics city every day, determine the proportionality that commercial power, commercial power are different with residential electricity consumption, analyze the daily load amount of every type load, Daily treatment cost and day minimum load three kinds of data, and draw its relation curve changed with above-mentioned influence factor respectively;

Step 2: carry out many types of compound: according to industry, business and the resident load scale-up factor of statistics, and the relation curve of every type load that step one is tried to achieve carries out compound, analyzes the many types of non-linear composite model family curve drawing whole load;

Step 3: carry out the modeling of polynary many types of composite non-linear regression model: is typical industry, business and residential electricity consumption family curve due to step one employing, also there is difference between the use electrical characteristics of urban whole and these residential electricity consumption characteristics, the former is the summation of industry, business and residential electricity consumption; Therefore need to re-start estimation to the parameter of the nonlinear function in the many types of composite model of step 2 foundation, obtain polynary many types of composite non-linear regression model more accurately;

Step 4: carry out Modifying model: the polynary many types of composite non-linear regression model that step 3 is obtained, adopt the test of hypothesis in mathematical statistics, by the load data sample of surveying, calculate the value of model prediction, and test according to the characteristic analyzing each type load obtained in advance, make the judgement of refusal or this model of acceptance hypothesis, thus obtain the quantitative data about this model credibility;

Urban power load short-term forecasting is completed by above-mentioned steps.

In described step one: comprising by electrical characteristics of urban whole affects larger temperature, wind speed, weather, date characteristic to power load.

In described step one: adopt polynomial curve fitting method, ask for the typical incidence relation between urban industry load and temperature, specify its relationship characteristic, in like manner obtain successively $F_I\left(T\right)=\underset{i=1}{\overset{NTI}{\Σ}}{a}_i^{TI}{f}_I\left(T^i\right),F_I\left(W\right)=\underset{i=1}{\overset{NWI}{\Σ}}{a}_i^{WI}{f}_I\left(W^i\right),$ $F_I\left(V\right)=\underset{i=1}{\overset{NVI}{\Σ}}{a}_i^{VI}{f}_I\left(V^i\right),F_I\left(D\right)=\underset{i=1}{\overset{NDI}{\Σ}}{a}_i^{DI}{f}_I\left(D^i\right)$ Other typical characteristics of isoperimetric load..

In described step 2: carry out compound and obtain the many types of compositive relation formula of following city load:

P＝K _I(F _I(T)+F _I(W)+F _I(V)+F _I(D))

+K _C(F _C(T)+F _C(W)+F _C(V)+F _C(D))

+K _A(F _A(T)+F _A(W)+F _A(V)+F _A(D))

In formula, K _i, K _cand K _abe respectively industry, business and residential electricity consumption scale-up factor.

In described step 3: in the Multiple Non Linear Regression parameter estimation of urban power load, measurement amount refers to the historical load of actual acquisition, state variable refers to the parametric variable relevant with temperature, weather, wind speed and date feature, and the ins and outs becoming power least square method comprise:

Between measurement amount and state variable, the non-linear measurement equation of relation can be expressed as:

z＝h(x)+v

Wherein, z is measurement amount, and x is quantity of state, the non-linear measurement function of h (x) for representing with quantity of state, v is error in measurement, and it is 0 that general supposition error in measurement v obeys average, standard deviation is the normal distribution of σ, and is separate between each measurement amount;

By v=z-h (x) obtain becoming power least-squares estimation objective function into:

minJ(x)＝[z-h(x)] ^TW _w[z-h(x)]

Be expressed as with scalar:

$\min J\left(x\right)=\underset{i=1}{\overset{m}{\Σ}}{w}_i{(z_i-h_i(x\left)\right)}^2$

In formula, m represents measurement amount number, W _wfor measuring weight matrix, w _ibe the weights of i-th measurement amount;

Objective function equals the weighted sum of squares of each measurement actual value and theory calculate value difference, and state estimation target makes J (x) minimum, and the state making J (x) minimum is exactly required state estimation.Objective function is minimized, the estimator of system state can be solved objective function is by iterative, and its iterative equation is:

△x ^(k)＝G ^-1(x ^(k))H ^T(x ^(k))W _w[z-h(x ^(k))]

x ^(k+1)＝x ^(k)+△x ^(k)

Wherein, x ^(k), x ^(k+1)represent the quantity of state after kth time and k+1 iteration respectively, for measuring Jacobian matrix; G (x ^(k))=H ^t(x ^(k)) W _wh (x ^(k)), be called information matrix;

Weight function selects stronger Robustness least squares, has again the Fair function of the valuation of greater efficiency, and the weight factor of Fair distribution weight function is:

$\mathrm{\ω}\left(v\right)=\frac{2}{1+\frac{\left|v\right|}{k\mathrm{\σ}}}$

In formula, v represents remaining difference, and σ represents the standard deviation of measurement data, and according to Normal Distribution Theory, the probability of error outside ± 1.5 σ is only 0.13;

The initial weight becoming power the least square estimation adopts the weights of basic weighted least-squares method, from second time iteration, obtains weight factor by the residual error produced after each iteration, is remodified the weights of each measurement amount by weight factor.By following formula, weights are modified:

$w_i^{(k+1)}=w_i^{\left(k\right)}*{\mathrm{\ω}}_i^{\left(k\right)}$

In formula, be respectively the weights of i-th measurement amount when k and k+1 iteration, for the weight factor of i-th measurement amount after kth time iteration.

The non-linear regression method of a kind of urban power load short-term forecasting of the present invention, technique effect is as follows:

1), multivariate regression model is adopted the factor affecting load variations all can be contributed in forecast model; Adopt nonlinear regression model (NLRM) effectively can avoid the deficiency of linear regression model (LRM), thus improve the precision of prediction.

2), the method is applicable to the urban power load short-term forecasting of China different regions, different economy level of development.

3), correlation analysis is carried out to historical load sample and influence factor thereof, historical load data are angularly carried out complex correlation analysis according to date type (working day, festivals or holidays etc.), meteorological condition (temperature, wind speed etc.), obtain the sensitivity of each influence factor to historical load, improve the sensitivity of load prediction.

4), with the Nanchang electrical network Power system load data of 2013 for object, adopt load forecasting method of the present invention to test, its precision of prediction reaches more than 95%, and prediction effect is good, and the software of its exploitation is applied at Nanchang electrical network.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention.

Fig. 2 is the temperature variant curve map of residential electricity consumption daily load amount of the present invention.

Embodiment

A kind of non-linear regression method of urban power load short-term forecasting, as shown in Figure 1, first the present invention considers the feature of urban power load, is divided into industrial load, Commercial Load and resident load three class, extracts the incidence relation between each type load and influence factor respectively.Secondly, on influence factor variable, emphasis considers temperature, wind speed, weather, date feature 4 factors larger on power load impact, follow-uply can expand, and wherein date feature refers to that forecast date is working day, weekend or great festivals or holidays.Multiple nonlinear regression method is finally utilized to ask for forecast model and parameter thereof.Specifically, urban electric power short-term load forecasting method comprises following step.

Step one: load Analysis:

The ratio of the city of China different regions, different economy level of development, commercial power, commercial power and residential electricity consumption is different.Wherein industrial load mainly comprises the electricity consumptions such as various extractive industry, Metallurgy Industry and manufacturing industry, accounts for whole electricity need load large percentage, and Commercial Load mainly refers to the electricity consumptions such as various shop, catering trade, and appliance load mainly contains the electrical equipment loads such as home lighting.The present invention analyzes the characteristic of different type in city electric load respectively according to the level shown in table 1, F _i(T) represent the daily load amount of industrial load, Daily treatment cost and the temperature variant funtcional relationship of day minimum load, other variable is similar with it, repeats no more.

The table 1 urban power load level of analysis

	Temperature (T)	Weather (W)	Wind speed (V)	Date feature (D)
					Industrial load (I)	F I(T)	F I(W)	F I(V)	F I(D)
Commercial Load (C)	F C(T)	F C(W)	F C(V)	F C(D)
					Resident load (A)	F A(T)	F A(W)	F A(V)	F A(D)

City has the advantage of Industry Agglomeration, for the industrial load in city, according to urban industries planning, carries out labor, obtain the industry of different cities to the typical incidence relation of temperature, weather, wind speed and date feature to the history week load of industrial park.Temperature refers to history temperature hourly every day; Weather variable refers to fine, cloudy, the discrete variable such as light rain, moderate rain, heavy rain, snow; Wind speed refers to the discrete variable of wind speed scale; Date feature then refers to the discrete variable of working day, weekend and festivals or holidays.Wherein, the quantized value of weather variable and date feature is expressed as follows:

Table 2 weather and date feature quantized value

The characteristic F of industry days load is affected with temperature _i(T) be example, consider many days accumulation associate features of temperature factor, suppose F _i(T) be following form:

$F_I\left(T\right)=\underset{i=1}{\overset{NTI}{\Σ}}{a}_i^{TI}{f}_I\left(T^i\right)$

In formula: NTI represents the number of days stronger with temperature dependence on the same day, value is 1 to 4, totally 4 points; represent that every day in the middle of one week, temperature factor affected weights to industrial load; T ⁱrepresent the medial temperature of this area's every day; f _i(T ⁱ) represent the pattern function that i-th day industrial load changes with temperature factor, can be linear function, parabolic function and power function etc., or many kinds of function be comprehensive.Namely the funtcional relationship between industrial load and temperature.

Adopt polynomial curve fitting method, ask for the typical incidence relation between urban industry load and temperature, specify its relationship characteristic.In like manner obtain successively $F_I\left(W\right)=\underset{i=1}{\overset{\mathrm{NWI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{WI}}{f}_I\left(W^i\right),F_I\left(V\right)=\underset{i=1}{\overset{\mathrm{NVI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{VI}}{f}_I\left(V^i\right)$ With $F_I\left(D\right)=\underset{i=1}{\overset{\mathrm{NDI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{DI}}{f}_I\left(D^i\right)$ Other typical characteristics of isoperimetric load.Wherein, $F_I\left(W\right)=\underset{i=1}{\overset{\mathrm{NWI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{WI}}{f}_I\left(W^i\right),F_I\left(V\right)=\underset{i=1}{\overset{\mathrm{NVI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{VI}}{f}_I\left(V^i\right)$ With $F_I\left(D\right)=\underset{i=1}{\overset{\mathrm{NDI}}{\mathrm{\Σ}}}{a}_i^{\mathrm{DI}}{f}_I\left(D^i\right)$ Represent weather W, wind speed V, date D tri-kinds of factors respectively to the influencing characteristic of urban industry load.

For commercial power and residential electricity consumption, also adopt the historical data of typical commercial accumulation area and resident's accumulation area ask for as stated above daily load amount, Daily treatment cost and day minimum load characteristic.

Step 2: many types of compound:

For the electric load of whole city, according to every type load typical load family curve that the industry added up, business and resident load scale-up factor and previous step are tried to achieve.Carry out compound and obtain the many types of compositive relation formula of following city load:

P＝K _I(F _I(T)+F _I(W)+F _I(V)+F _I(D))

+K _C(F _C(T)+F _C(W)+F _C(V)+F _C(D))

+K _A(F _A(T)+F _A(W)+F _A(V)+F _A(D))

In formula, K _i, K _cand K _abe respectively industry, business and residential electricity consumption scale-up factor.

F _i(T) for temperature affects the characteristic function of commercial power daily load amount, F _i(W) be the characteristic function of weather effect commercial power daily load amount, F _i(V) be the characteristic function of air speed influence commercial power daily load amount, F _i(D) for the date affects the characteristic function of commercial power daily load amount;

F _c(T) for temperature affects the characteristic function of commercial power daily load amount, F _c(W) be the characteristic function of weather effect commercial power daily load amount, F _c(V) be the characteristic function of air speed influence commercial power daily load amount, F _c(D) for the date affects the characteristic function of commercial power daily load amount;

F _a(T) for temperature affects the characteristic function of residential electricity consumption daily load amount, F _a(W) be the characteristic function of weather effect residential electricity consumption daily load amount, F _a(V) be the characteristic function of air speed influence residential electricity consumption daily load amount, F _a(D) for the date affects the characteristic function of residential electricity consumption daily load amount.

Step 3: polynary many types of composite non-linear regression model modeling:

Due to step one employing is typical industry, business and residential electricity consumption family curve, also there is difference between the use electrical characteristics of urban whole and these characteristics, therefore need to re-start estimation to the parameter of the nonlinear function in the many types of composite model of step 2 foundation.Specifically, in industrial load characteristic, parameter comprises: the parameter relevant with temperature the parameter relevant with weather the parameter relevant with wind speed and the parameter relevant with date feature the same industrial load of parameter in business and resident load characteristic.The parameter estimation of Multiple Non Linear Regression adopts and becomes power least square method, and the method is weighted correction to basic least square model, has good Robustness least squares, can make up and determine the deficiency of Weights-selected Algorithm in robust.

In the Multiple Non Linear Regression parameter estimation of urban power load, measurement amount refers to the historical load of actual acquisition, and state variable refers to the parametric variable relevant with temperature, weather, wind speed and date feature.The ins and outs becoming power least square method comprise:

Between measurement amount and state variable, the non-linear measurement equation of relation can be expressed as:

z＝h(x)+v

Wherein, z is measurement amount, and x is quantity of state, the non-linear measurement function of h (x) for representing with quantity of state, v is error in measurement, and it is 0 that general supposition error in measurement v obeys average, standard deviation is the normal distribution of σ, and is separate between each measurement amount.

By v=z-h (x) obtain becoming power least-squares estimation objective function into:

minJ(x)＝[z-h(x)] ^TW _w[z-h(x)]

Be expressed as with scalar:

$\min J\left(x\right)=\underset{i=1}{\overset{m}{\Σ}}{w}_i{(z_i-h_i(x\left)\right)}^2$

In formula, m represents measurement amount number, W _wfor measuring weight matrix, w _ibe the weights of i-th measurement amount.

Objective function equals the weighted sum of squares of each measurement actual value and theory calculate value difference, and state estimation target makes J (x) minimum, and the state making J (x) minimum is exactly required state estimation.Objective function is minimized, the estimator of system state can be solved objective function is by iterative, and its iterative equation is:

△x ^(k)＝G ^-1(x ^(k))H ^T(x ^(k))W _w[z-h(x ^(k))]

x ^(k+1)＝x ^(k)+△x ^(k)

Wherein, x ^(k), x ^(k+1)represent the quantity of state after kth time and k+1 iteration respectively, for measuring Jacobian matrix; G (x ^(k))=H ^t(x ^(k)) W _wh (x ^(k)), be called information matrix.

Weight function selects stronger Robustness least squares, has again the Fair function of the valuation of greater efficiency, and the weight factor of Fair distribution weight function is:

$\mathrm{\ω}\left(v\right)=\frac{2}{1+\frac{\left|v\right|}{k\mathrm{\σ}}}$

In formula, v represents remaining difference, and σ represents the standard deviation of measurement data.According to Normal Distribution Theory, the probability of error outside ± 1.5 σ is only 0.13.

The initial weight becoming power the least square estimation adopts the weights of basic weighted least-squares method, from second time iteration, obtains weight factor by the residual error produced after each iteration, is remodified the weights of each measurement amount by weight factor.By following formula, weights are modified:

$w_i^{(k+1)}=w_i^{\left(k\right)}*{\mathrm{\ω}}_i^{\left(k\right)}$

In formula, be respectively the weights of i-th measurement amount when k and k+1 iteration, for the weight factor of i-th measurement amount after kth time iteration.

Step 4: Modifying model:

In mathematical statistics, test of hypothesis refers to the significant difference inspection between sample statistic and the population parameter of hypothesis, use small probability principle, determine the limits criteria of judgement in advance, if the probability that the null hypothesis of proposition calculates is less than this standard, just refuse null hypothesis, be greater than this standard and then accept null hypothesis.For the polynary many types of composite non-linear regression model that step 3 obtains, by the load data sample of surveying, calculate the value of model prediction, and test according to level of significance given in advance, make the judgement of refusal or this model of acceptance hypothesis, thus obtain the quantitative data about this model credibility.

Be illustrated in figure 2 the temperature variant curve of residential electricity consumption daily load amount in January 1 to March 26 in certain city first quarter, it is F that regretional analysis obtains model _a(T)=10.6466 × T ²-438.8934 × T+13457.

In like manner, it is similar to residential electricity consumption that analysis obtains the temperature variant model of commercial power daily load amount, is also quadratic function: F _c(T)=8.1563 × T ²-469.5120 × T+9132; And commercial power daily load amount and temperature dependency are only linear, model is: F _i(T)=-1.0154 × T+25189.

Regretional analysis three type load is identical with said method with the model of other weather, wind speed and date feature three kinds of influence factors, repeats no more.

Step 2 carries out daily load characteristic compound tense, and the ratio of industrial load, Commercial Load and resident load is respectively 0.54, and 0.19,0.27; Compound obtains this city day power load

P＝7×(1.3×T+0.3×W+0.005×V+0.07×D) ²-

290.5×(1.3×T+0.3×W+0.005×V+0.07×D)+44300

It can thus be appreciated that the temperature variant parameter of this city daily load is 7,1.3 and 290.5, is designated as a, b and c respectively; Daily load is 7,0.3 and 290.5 with the parameter of Changes in weather, is designated as a, d and c respectively; The parameter that daily load changes with wind speed is 7,0.005 and 290.5, is designated as a, e and c respectively; The parameter that daily load changes with date feature is 7,0.07 and 290.5, is designated as a, f and c respectively; Constant coefficient is designated as g.Therefore this city day power load model can be expressed as:

P＝a×(b×T+d×W+e×V+f×D) ²-

c×(b×T+d×W+e×V+f×D)+g

Using a, b, c, d, e, f and g as undetermined parameter, the change power least square method utilizing step 3 to carry carries out matching, and determine accurate coefficient further, then obtaining daily load model is

P＝8.12×(1.202×T+0.312×W+0.005×V+0.07×D) ²-

309.7×(1.202×T+0.312×W+0.005×V+0.07×D)+44854.62

Using this March 27 city first quarter to March 31 as reservation data, above-mentioned model is tested.Table 3 gives the data of reservation day and predicts the outcome.

Table 3 retention date data and predicting the outcome

Can find out that the relative error between actual daily load amount and prediction daily load amount meets the requirements substantially, the forecast model that the method demonstrating above-mentioned load prediction obtains has certain using value.In the prediction work of reality, also should interval certain cycle, such as 2 days, choose the historical load data in this city, utilize said method to recalculate forecast model, ensure the real-time effectiveness of model with this.

In short-term load forecasting, according to said method flow process dope Daily treatment cost and day minimum load so that traffic department prepares the load condition learning next day, arrange time daily output plan of each generating plant.

Claims (5)

1. a non-linear regression method for urban power load short-term forecasting, is characterized in that comprising the following steps:

Step one: carry out load Analysis: the load of quantitative statistics city every day, determine the proportionality that commercial power, commercial power are different with residential electricity consumption, analyze the daily load amount of every type load, Daily treatment cost and day minimum load three kinds of data, and draw its relation curve changed with above-mentioned influence factor respectively;

Step 2: carry out many types of compound: according to industry, business and the resident load scale-up factor of statistics, and the relation curve of every type load that step one is tried to achieve carries out compound, analyzes the many types of non-linear composite model family curve drawing whole load;

Step 3: carry out the modeling of polynary many types of composite non-linear regression model: is typical industry, business and residential electricity consumption family curve due to step one employing, also there is difference between the use electrical characteristics of urban whole and these residential electricity consumption characteristics, the former is the summation of industry, business and residential electricity consumption; Therefore need to re-start estimation to the parameter of the nonlinear function in the many types of composite model of step 2 foundation, obtain polynary many types of composite non-linear regression model more accurately;

Step 4: carry out Modifying model: the polynary many types of composite non-linear regression model that step 3 is obtained, adopt the test of hypothesis in mathematical statistics, by the load data sample of surveying, calculate the value of model prediction, and test according to the characteristic analyzing each type load obtained in advance, make the judgement of refusal or this model of acceptance hypothesis, thus obtain the quantitative data about this model credibility;

Urban power load short-term forecasting is completed by above-mentioned steps.

2. the non-linear regression method of a kind of urban power load short-term forecasting according to claim 1, is characterized in that, in described step one: comprising by electrical characteristics of urban whole affects larger temperature, wind speed, weather, date characteristic to power load.

3. the non-linear regression method of a kind of urban power load short-term forecasting according to claim 1, it is characterized in that, in described step one: adopt polynomial curve fitting method, ask for the typical incidence relation between urban industry load and temperature, specify its relationship characteristic, in like manner obtain successively $F_I\left(T\right)=\underset{i=1}{\overset{NTI}{\Σ}}{a}_i^{TI}{f}_I\left(T^i\right),F_I\left(W\right)=\underset{i=1}{\overset{NWI}{\Σ}}{a}_i^{WI}{f}_I\left(W^i\right),$ $F_I\left(V\right)=\underset{i=1}{\overset{NVI}{\Σ}}{a}_i^{VI}{f}_I\left(V^i\right),F_I\left(D\right)=\underset{i=1}{\overset{NDI}{\Σ}}{a}_i^{DI}{f}_I\left(D^i\right)$ Other typical characteristics of isoperimetric load..

4. the non-linear regression method of a kind of urban power load short-term forecasting according to claim 1, is characterized in that, in described step 2: carry out compound and obtain the many types of compositive relation formula of following city load:

P=K _I(F _I(T)+F _I(W)+F _I(V)+F _I(D))

+K _C(F _C(T)+F _C(W)+F _C(V)+F _C(D))

+K _A(F _A(T)+F _A(W)+F _A(V)+F _A(D))

In formula, K _i, K _cand K _abe respectively industry, business and residential electricity consumption scale-up factor.

5. the non-linear regression method of a kind of urban power load short-term forecasting according to claim 1, is characterized in that, in described step 3:

In the Multiple Non Linear Regression parameter estimation of urban power load, measurement amount refers to the historical load of actual acquisition, and state variable refers to the parametric variable relevant with temperature, weather, wind speed and date feature, and the ins and outs becoming power least square method comprise:

Between measurement amount and state variable, the non-linear measurement equation of relation can be expressed as:

z=h(x)+v

Wherein, z is measurement amount, and x is quantity of state, the non-linear measurement function of h (x) for representing with quantity of state, v is error in measurement, and it is 0 that general supposition error in measurement v obeys average, standard deviation is the normal distribution of σ, and is separate between each measurement amount;

By v=z-h (x) obtain becoming power least-squares estimation objective function into:

minJ(x)=[z-h(x)] ^TW _w[z-h(x)]

Be expressed as with scalar:

$\min J\left(x\right)=\underset{i=1}{\overset{m}{\Σ}}{w}_i{(z_i-h_i(x\left)\right)}^2$

In formula, m represents measurement amount number, W _wfor measuring weight matrix, w _ibe the weights of i-th measurement amount;

Objective function equals the weighted sum of squares of each measurement actual value and theory calculate value difference, and state estimation target makes J (x) minimum, and the state making J (x) minimum is exactly required state estimation; Objective function is minimized, the estimator of system state can be solved objective function is by iterative, and its iterative equation is:

△x ^(k)=G ^-1(x ^(k))H ^T(x ^(k))W _w[z-h(x ^(k))]

x ^(k+1)=x ^(k)+△x ^(k)

Wherein, x ^(k), x ^(k+1)represent the quantity of state after kth time and k+1 iteration respectively, for measuring Jacobian matrix; G (x ^(k))=H ^t(x ^(k)) W _wh (x ^(k)), be called information matrix;

Weight function selects stronger Robustness least squares, has again the Fair function of the valuation of greater efficiency, and the weight factor of Fair distribution weight function is:

$\mathrm{\ω}\left(v\right)=\frac{2}{1+\frac{\left|v\right|}{k\mathrm{\σ}}}$

In formula, v represents remaining difference, and σ represents the standard deviation of measurement data, and according to Normal Distribution Theory, the probability of error outside ± 1.5 σ is only 0.13;

The initial weight becoming power the least square estimation adopts the weights of basic weighted least-squares method, from second time iteration, obtain weight factor by the residual error produced after each iteration, remodified the weights of each measurement amount by weight factor, by following formula, weights are modified:

$w_i^{(k+1)}=w_i^{\left(k\right)}*{\mathrm{\ω}}_i^{\left(k\right)}$

In formula, be respectively the weights of i-th measurement amount when k and k+1 iteration, for the weight factor of i-th measurement amount after kth time iteration.