CN104463683A - Long-term load prediction device and method in power grid with multiple sources - Google Patents

Abstract

The invention belongs to the field of long-term load prediction in an electric power system, and particularly relates to a long-term load prediction device and method in a power grid with multiple sources. A double-CPU unit, a GPRS communication unit, a storage unit, a human-computer interaction unit and a power unit are arranged in a device box, wherein the double-CPU unit is connected with the GPRS communication unit, the human-computer interaction unit, the power unit and the storage unit, and power supply is achieved through the power unit. Through the device and method, the relation between various factors and a prediction result can be objectively reflected, and a high-dimensional characteristic space is built; a low-dimensional characteristic having best distinguishing capacity is extracted from the high-dimensional characteristic space in a combined and optimized mode, excessive learning can be avoided to the maximum extent, and algorithm efficiency is improved; a load prediction process can be faster and more effective, long-term load prediction accuracy degree in the power grid with the multiple sources can be improved easily, and therefore match of power generation areas in the power grid can be adjusted, dispatch of electric network power flow is performed, and the overall power generation level of the power grid in the area can meet the overall requirement of loads in the area in future.

Description

A kind of containing multi-source electrical network Mid-long term load forecasting device and method

Technical field

The invention belongs to power system mid-long term load forecasting field, particularly one is containing multi-source electrical network Mid-long term load forecasting device and method.

Background technology

At present, China is in and greatly develops the new forms of energy stage, rack is in the fast-developing construction period, network load total demand is the element task of urban power network planning, which determine the power supply capacity of future city to the demand of electric power and future city electrical network, the height of its precision directly affects the quality of Electric Power Network Planning quality.Therefore, the prediction of workload demand total amount becomes one of key link in power grid construction.

The many employings of prior art predict the workload demand total amount of electrical network in a certain period based on the Forecasting Methodology of learning training, but, the problem that currently used method ubiquity forecasting inaccuracy is true.In this class Forecasting Methodology, core link is the training of learner, in order to ensure the computing velocity predicted, and the method construct learner adopting weak typing more, and the judgement of mistake may be there is for the classification of sample characteristics in Weak Classifier, make the mistake study.Because Mid-long Term Load demand is subject to power system operating mode, power supply composition, load type, regional economic development degree, administrative planning etc. many factors, its factor feature generally lacks conspicuousness with regular, cause current many Forecasting Methodology mistake study situations relatively more serious, Mid-long Term Load demand forecast problem can not be applicable to exactly.

Up to now, Mid-long term load forecasting problem is mainly adopted to the methods such as intelligent learning algorithm, gray system theory, expert system, support vector machine, not for Forecasting Methodology and the autonomous device of considering a large amount of new forms of energy access electrical network, only a large amount of historical data is carried out indifferenceization input to calculate, analysis selection is not carried out to Mid-long term load forecasting influence factor in appointment electrical network (containing multi-source electrical network), resultant error is large, with a low credibility.Simultaneously, needed for Mid-long term load forecasting calculates, basic data amount is large, it is heavy that Electric Power Network Planning department obtains the workload that predicts the outcome, and therefore designs for considering that the work containing multi-source electrical network Mid-long term load forecasting method and apparatus of a large amount of new forms of energy access electrical network seems of crucial importance.

Summary of the invention

For the deficiencies in the prior art, the object of this invention is to provide a kind of containing multi-source electrical network Mid-long term load forecasting device and method.

Technical scheme of the present invention is achieved in that

A kind of containing multi-source electrical network Mid-long term load forecasting device, be in device case, be provided with dual processors unit, GPRS communication unit, storage unit, man-machine interaction unit and power supply unit; Wherein, dual processors unit is connected with GPRS communication unit, man-machine interaction unit, power supply unit, storage unit respectively; And powered by power supply unit.

Described man-machine interaction unit: for obtaining regional economy class data and weather data, wherein, economic class data comprise: population, secondary industry total value, gross national product (GNP), inhabitant's consumption level and average electricity price; Weather data comprises: annual sunshine-duration, annual mean wind speed;

Described dual processors unit: for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors of Urban Annual Electrical Power Consumption total amount carry out affecting intensive analysis, excavate the relation between different year, dissimilar data, the annual workload demand total value of somewhere electrical network is then predicted;

Described GPRS communication unit: new forms of energy speculate operation amount, new forms of energy increase operation amount, then load peak, then load valley newly, these service datas of Urban Annual Electrical Power Consumption total amount for obtaining from dispatching center, and input to dual processors unit and calculate; Described dispatching center, i.e. grid dispatching center data server, for prediction unit the actual operation amount of new forms of energy is provided, new forms of energy increase operation amount, then load peak, then load valley newly, these operation of power networks historical datas of Urban Annual Electrical Power Consumption total amount, and obtain this area's load prediction results data;

Described storage unit: carrying out summing up editor for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors historical datas of Urban Annual Electrical Power Consumption total amount and the regional load data that predict the outcome, making data sheet.

Described dual processors unit comprises ARM embedded microcontroller and dsp processor; Described GPRS communication unit adopts two-way RS485 circuit, using SP485R chip as transceiver; Also be provided with filtering dividing potential drop anti-jamming circuit in described GPRS communication unit, described filtering dividing potential drop anti-jamming circuit comprises further: by diode D ₁, diode D ₂, diode D ₃, diode D ₄, form anti-interference rectifier circuit; Diode D ₁with diode D ₂tie point, with diode D ₃with diode D ₄tie point between be connected in parallel filter inductance L ₁, filter capacitor C ₃; Divider resistance R ₇, divider resistance R ₈with electric capacity of voltage regulation C ₁connect in angle-style, and divider resistance R ₈with electric capacity of voltage regulation C ₁tie point place and diode D ₂, diode D4 tie point place be connected; Divider resistance R ₇with electric capacity of voltage regulation C ₁tie point place series diode D ₅with common mode inductance wave filter ZJYS51; Diode D ₅and contact resistance C between ZJYS51 ₄one end, resistance C ₄other end ground connection; Divider resistance R ₇, divider resistance R ₈tie point place be connected in series divider resistance R ₉, simultaneously at divider resistance R ₉upper parallel voltage-stabilizing electric capacity C ₂.

The input/output terminal of described ARM embedded microcontroller connects man-machine interaction unit, and its output terminal connects dsp processor; Described man-machine interaction unit comprises button and LCDs; Described button connects the first input/output terminal of ARM embedded microcontroller, and described LCDs connects the second input/output terminal of ARM embedded microcontroller.

A kind of containing multi-source electrical network Mid-long term load forecasting method, utilize the one described in claim 1 to realize containing multi-source electrical network Mid-long term load forecasting device, comprise the steps:

Step 1, by year chooses region factors historical data and network load demand measured value as data sample;

Step 2: the data in integrating step 1, by excavating relation between different year, dissimilar data, find the sample weights of region factors sample for workload demand measured value, for determining the influence degree of the region factors of different year for network load demand;

Step 3: adopt the prediction algorithm based on linear mapping, for predicting the annual workload demand total value of somewhere electrical network then;

Step 4: according to the regional factor of actual area electrical network, utilize the regression equation that step 3 is determined, calculate the workload demand total value of this regional power grid, utilize this composite demand total value, adjust the cooperation of each power generation region in this electrical network, carry out electric network swim scheduling, enable the overall power generation level of this area's electrical network support the overall needs of following regional load.

Described step 1, by year chooses region factors historical data and network load demand measured value as data sample; First, the time will predicting network load demand is determined; Then, after determining this time, the historical data of this regional factor and the network load demand measured value corresponding with this year, this area in each time before being chosen at this time, at least 5 years; Described region factors historical data comprises following 12 item number certificates: total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy increase operation amount newly, the annual sunshine-duration, annual mean wind speed, load peak then, then load valley, Urban Annual Electrical Power Consumption total amount.

Described step 2: 12 item number certificates in integrating step 1, by excavating relation between different year, dissimilar data, find the sample weights of region factors sample for workload demand measured value, for determining the influence degree of the region factors of different year for network load demand; Wherein to the calculating of weight, concrete steps are as follows:

(1) build hierarchical structure, described hierarchical structure forms by three layers, is specially:

Destination layer: be the network load demand measured value that each time before this area's network load to be predicted demand time, at least 5 years is corresponding; Different year is set objectives layer respectively;

Rule layer: based on the time of destination layer, before getting this time, time of at least 3 years, forms the sample weights time; Each destination layer determines the respective sample weights time, forms respective rule layer respectively;

Solution layer: the time determined according to rule layer, calculates the changing sensitivity of each region factors historical data under the change of same load demand measured value under this time; Each solution layer determines different regions factor changing sensitivity separately, forms solution layer respectively;

(2) standardization processing is carried out to the data in different levels, is specially:

Destination layer: destination layer represents the problem that will solve, and its data do not need standardization processing;

Rule layer: adopt scale to quantize criterion and carry out standardization processing;

Solution layer: first, carries out standardization processing according to the determined different year of rule layer, i.e. normalized respectively for the every of solution layer;

Then, the sensitivity of all samples for load variations is determined;

The sensitivity of described load variations reflects, the change of workload demand value and the relation between the adjacent time between sample changed between the adjacent time;

(3) by development of judgment matrix, ask for the weight of each sample of solution layer for the corresponding rule layer time, and rule layer each time is for the weight of destination layer workload demand;

(4) to the synthetic weights severe W of solution layer for destination layer _icalculate, namely determine that every sample in destination layer time affects weights for workload demand value, formula is:

W _i＝Σ _jp _jv _ij

In formula, p _jfor the correlativity between rule layer time and destination layer workload demand value, and there is j=1 ... n, wherein n>=3; v _ijfor the correlativity between solution layer and rule layer, and there is i=1 ... m, wherein, m=12.

Described step 3: adopt the prediction algorithm based on linear mapping, for predicting the annual workload demand total value of somewhere electrical network then, is specially:

Step 3.1: treat the region factors sample predicting the regional network load demand time, and before this year, the region factors sample in some times, workload demand sample are normalized; And to the region factors sample evidence different year i after normalized, that determines with step 2 affects weights W _ibe multiplied process;

Step 3.2: structure original sample collection, as the input based on linear mapping prediction algorithm;

Be input amendment by 12 region factors samples in step (3.1) and full regional load demand is as output quantity set up classified sample set S; As follows:

$S=\left\{(x_i^0,y_i^0)\right|x_i^0\∈R^L,y_i^0\∈R,i=\mathrm{1,2},...,N\}$

In formula, N represents the quantity of input amendment, the representative sample time in sample set S, and L represents initial sample number 12;

The classified sample set obtained in step 3.2 is converted to two class classified sample sets, by described two class sample set combinations, obtains instructing the original sample collection D based on the prediction algorithm of linear mapping; As follows:

D＝{(x _i,y _i)|x _i∈R ^L+1,y _i∈{-1,+1},i＝1,2,...,2N}

In formula, x _irepresent 12 samples and value the formed vector of workload demand measured value after translation in step 1; I represents sample sequence number, and 2N represents the quantity of sample; y _ifor the desired value of the original sample collection of structure, actual value is 1 or-1;

Step 3.3: the sample set D utilizing structure in step 3.2, calculates initial sample weights vector d _{1, i}, according to y _ivalue is 1 or-1 sample is considered as two classes, for each sample in each class, carries out weight average distribution according to total sample number amount, and ensures that overall all sample weights sums are 1;

Set up sample average m _t,k, k=1 or-1, t represent iterations; Corresponding two kinds of representative sample average is categorized as two; Each sample average relies on method of weighted mean to expect to calculate sample by meeting a class sample interior, and obtains divided by the total weight in this classification, and because sample is through normalized, in sample average, any element should not more than 1;

Step 3.4: set up the linear mapping factor;

To utilize sample to be 1 be-1 with sample, and sample average constructs sample standard deviation value difference, and constructs between-class scatter and within-class scatter in L+1 dimension space after the conversion, and two dispersions described in utilization, determine the linear mapping factor; Between-class scatter and within-cluster variance computing formula are:

$S_{t,k}=\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}{x}_i(x_i-m_{t,k}){(x_i-m_{t,k})}^T/\left(\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}e\right),S_{t,W}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}(x_i-m_{t,k}){(x_i-m_{t,k})}^T$

In formula, S _t,kfor the inter _ class relationship of sample, S _t,Wfor the within-cluster variance of sample; d _t,ifor the sample weights vector of current iteration layer t, its value relies on the right value update in each iteration; E representation dimension and sample x _iidentical vector, its inner element is 1, m _t,kfor two class sample averages in current iteration layer; I represents sample sequence number;

The linear mapping factor computing method are the business of inter _ class relationship and within-cluster variance; Original sample can be mapped to new sample space by the linear mapping factor, and sample meets inter _ class relationship and the maximization of within-cluster variance ratio within this space;

Step 3.5: set up Weak Classifier;

Weak Classifier by best projection map construction current iteration layer: when the sample in mapping space is greater than threshold values, be divided into a class, otherwise as another kind of, the Weak Classifier formula of foundation is:

h _t＝{h _t,l|l＝1...L+1}

$h_{t,l}\left(x_{i,l}\right)=\left\{\begin{array}{c}+1,{\left(w_t^{*}{x}_i\right)}_l>{\left({\mathrm{\θ}}_t\right)}_l\\ -1,\mathrm{others}\end{array}\right.$

In formula, h _tby being constructed Weak Classifier, x _i,lrepresent x _iin l element, for in l element, l=1 ... L+1 representative sample type, i represents sample sequence number, θ _tbe threshold values vector, determined by the classification error rate in previous iteration layer, (θ _t) _lit is its inner l element;

The computing formula of the vector of described threshold value is:

${\mathrm{\θ}}_t=\left\{\begin{array}{c}{w}_t^{*}{m}_t(1-{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}),{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}>{\mathrm{\ϵ}}_{t-1}^{(-1)}\\ w_t^{*}{m}_t(1+{\mathrm{\ϵ}}_{t-1}^{(-1)}),{\mathrm{\ϵ}}_{t-1}^{(-1)}>{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}\end{array}\right.$

In formula, m _tfor mean vector, dimension and sample x _iidentical, inner element is the average of the sample average in current iteration, i.e. m _{t, 1}with m _{t ,-1}average, for meeting y in previous iteration _ithe classification error rate of=1 sample, for meeting y in previous iteration _ithe classification error rate of=-1 sample, wherein, during first time iterative computation t=1, classification error rate is 0;

Step 3.6: determine classification error rate, formula is:

${\mathrm{\ϵ}}_t^{\left(1\right)}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}I(h_t{\left(x_i\right)}_l\≠1)$

${\mathrm{\ϵ}}_{t-1}^{(-1)}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}I(h_t{\left(x_i\right)}_l\≠-1)$

In formula, h _t(x _i) represent the classification results of sample, h _t(x _i) _lfor its inside l element, i represents sample sequence number, h _t(x _i) _l≠ 1 expression meets y _ii-th sample x of=1 _iin, l sample elements x _l,iby mis-classification in current iteration layer t, h _t(x _i) _l≠-1 implication is similar, and I () represents discriminant function, if the content in bracket is set up, then I=1, otherwise I=0, finally to the weight d of all misjudged samples _{t, l, i}respectively with y _i=1 or y _i=-1 carries out classification weighting, obtains two class classification error rates in current iteration layer;

Step 3.7: utilize the classification error rate in step (3.6) to determine parameter alpha of voting _t, formula is:

${\mathrm{\α}}_t=(1/2)\ln ((1-{\mathrm{\ϵ}}_t)/{\mathrm{\ϵ}}_t),{\mathrm{\ϵ}}_t={\mathrm{\ϵ}}_t^{\left(1\right)}+{\mathrm{\ϵ}}_t^{(-1)}$

In formula, ε _tfor the classification error rate of samples all in current iteration layer; Ballot parameter alpha _tas the t time iteration obtain sorter classification accuracy rate one weigh, when it act as and represents sorter in final weighting all layers, the weight of each layer sorter, the t time Iterative classification error rate is less, represents that the t time Iterative classification device is for totally more effective;

Step 3.8: upgrade weight vectors by classification error rate and ballot parameter, the weight vectors formula after renewal is as follows:

d _t+1,i＝d _t,i*exp(-α _ty _ih _t(x _i))

In formula, d _t,ifor original weight vectors, y _irepresent score value 1 or-1, its value corresponds to x _imiddle element; The inner element of exp function is vector, and exp take e as the truth of a matter, with each element in vector in bracket for index, carries out exponential function calculating; In formula, No. * represents vector point multiplication, and namely vectorial inner respective items is multiplied, and its result of calculation is the vector that dimension is identical, the d after namely upgrading _{t+1, i};

Weight in weight vectors should meet and adds and equal 1, therefore is normalized formula to weight vectors and is:

$d_{t+1,i}=\frac{1}{Z_t}{d}_{t+1,i}$

In formula, Zt equal weight vectors interior element adding and;

Repeated execution of steps 3.1 ~ step 3.8, until the training completing all samples;

Step 3.9: Weak Classifier in all iteration layers finally step 3.5 determined, the ballot parameter determined using step 3.7 is as respective weights, utilize the integrated all sorters of weighting, obtain regression equation, this regression equation describes the implicit function relation between sample data and target load requirements, it is input as 12 master datas in target time, exports the predicted value being area power grid workload demand data.

Determine that all samples for the computing method of the sensitivity of load variations are in solution layer in (2) of described step 2:

12 sampled data values of n before this time are designated as c _i,j, i=1 ... m, m=12, j=1 ... n, n>=3; Workload demand measured value is designated as a _j, j=1 ... n; Before this time, the quantification in rule layer time is designated as b _j, j=1 ... n; Correlativity between rule layer time and destination layer workload demand value is p _j, solution layer changing sensitivity is designated as q _ij, i=1 ... m, the correlativity between solution layer and rule layer is v _ij, for the time in each rule layer, calculate the every quantized value q in respective party pattern layer by following formula _ij, namely all samples are for the sensitivity of load variations:

$q_{\mathrm{ij}}=\frac{a_j-a_{j-1}}{c_{i,j}-c_{i,j-1}}$

In formula, q _ijimplication is: i-th sample in a jth time is q for the changing sensitivity of a jth time workload demand value _ij.

Ask for the weight of each sample of solution layer for the corresponding rule layer time described in (3) in described step 2, and rule layer each time is for the weight of destination layer workload demand, is specially:

Have 1 for rule layer judgment matrix A, the time quantity that dimension n is comprised by solution layer determines, wherein element is:

$A=\left[\begin{array}{ccc}\frac{b_1}{b_1}& ...& \frac{b_1}{b_n}\\ ...& & ...\\ \frac{b_n}{b_1}& ...& \frac{b_n}{b_n}\end{array}\right]$

Judgment matrix B for solution layer has n, and dimension m is determined by the sample size participated in step 1, for a jth matrix wherein element be:

$B_j=\left[\begin{array}{ccc}\frac{q_{1j}}{q_{1j}}& ...& \frac{q_{1j}}{q_{\mathrm{mj}}}\\ ...& & ...\\ \frac{q_{\mathrm{mj}}}{q_{1j}}& ...& \frac{q_{\mathrm{mj}}}{q_{\mathrm{mj}}}\end{array}\right]$

Calculate eigenvalue of maximum and the proper vector of each matrix, tried to achieve the weight vectors of disparity items by proper vector, weight vectors is designated as respectively:

p＝(p ₁,p ₂,...p _j,...p _n),j＝1..n

v＝(v _1j,v _2j,...,v _ij,...v _mj),i＝1..m

In formula, in p expressiveness layer, different year is for the weight of workload demand; Vj represents that different samples in the rule layer jth time are to the weight of workload demand.

Beneficial effect of the present invention: the present invention's one is containing multi-source electrical network Mid-long term load forecasting method and device, in the Mid-long term load forecasting of a large amount of distributed power source access electrical network calculates, analyze selection add new forms of energy factor and quantize (operation amount as actual in new forms of energy to it, new forms of energy increase operation amount newly, wind energy turbine set area annual mean wind speed, regional average sunshine in photovoltaic power plant year etc.), carry out calculating from the excellent Changeable weight that becomes on each factor affecting load variations, guarantee the relation objectively responding each factor Yu predict the outcome, set up high-dimensional feature space, and therefrom to extract and Combinatorial Optimization goes out to have most the low dimensional feature of discriminating power, train up data, maximization is avoided producing overlearning, improve efficiency of algorithm, make load prediction process more fast effectively, be beneficial to and improve containing the accuracy of multi-source electrical network Mid-long term load forecasting, and then adjust the cooperation of each power generation region in this electrical network, carry out electric network swim scheduling, the overall power generation level of this area's electrical network is enable to support the overall needs of following regional load.

Accompanying drawing explanation

Fig. 1 is that the general construction containing multi-source electrical network Mid-long term load forecasting device in embodiment of the present invention is always schemed;

Fig. 2 is dual-cpu structure hardware circuit principle figure in embodiment of the present invention;

Fig. 3 is communication unit GPRS hardware circuit principle figure in embodiment of the present invention;

Fig. 4 is filtering dividing potential drop anti-jamming circuit schematic diagram in embodiment of the present invention;

Fig. 5 is the interface circuit schematic diagram between embodiment of the present invention GPRS and RS485;

Fig. 6 is storage unit hardware circuit principle figure in embodiment of the present invention;

Fig. 7 is OLED interface circuit schematic diagram in man-machine interaction unit in embodiment of the present invention;

Fig. 8 is button hardware circuit principle figure in man-machine interaction unit in embodiment of the present invention;

Fig. 9 is containing multi-source electrical network Mid-long term load forecasting method flow diagram in embodiment of the present invention.

In figure: dual processors unit 1, GPRS communication unit 2, storage unit 3, man-machine interaction unit 4, ARM embedded microcontroller 5, dsp processor 6, two-way RS485 circuit 7, filtering dividing potential drop anti-jamming circuit 8, GPRS-DTU chip 9, LCDs 10, the interface circuit 11 of button ARM.

Embodiment

Below in conjunction with accompanying drawing, embodiments of the present invention are further described in detail.

The general structure containing multi-source electrical network Mid-long term load forecasting device in present embodiment as shown in Figure 1.Comprise: dual processors unit 1, GPRS communication unit 2, storage unit 3 and man-machine interaction unit 4.Specifically in device case, be provided with dual processors unit 1, GPRS communication unit 2, storage unit 3, man-machine interaction unit 4 and power supply unit; Wherein, dual processors unit is connected with GPRS communication unit, man-machine interaction unit, power supply unit, storage unit respectively; And powered by power supply unit, power supply unit adopts 12V storage battery power supply.

Described man-machine interaction unit 4: for obtaining regional economy class data and weather data, wherein, economic class data comprise: population, secondary industry total value, gross national product (GNP), inhabitant's consumption level and average electricity price; Weather data comprises: annual sunshine-duration, annual mean wind speed;

Described dual processors unit 1: for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors of Urban Annual Electrical Power Consumption total amount carry out affecting intensive analysis, excavate the relation between different year, dissimilar data, the annual workload demand total value of somewhere electrical network is then predicted;

Described GPRS communication unit 2: new forms of energy speculate operation amount, new forms of energy increase operation amount, then load peak, then load valley newly, these service datas of Urban Annual Electrical Power Consumption total amount for obtaining from dispatching center, and input to dual processors unit and calculate.Described dispatching center: i.e. grid dispatching center data server, for prediction unit the actual operation amount of new forms of energy is provided, new forms of energy increase operation amount, then load peak, then load valley newly, these operation of power networks historical datas of Urban Annual Electrical Power Consumption total amount, and obtain this area's load prediction results data.

Described storage unit 3: carrying out summing up editor for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors historical datas of Urban Annual Electrical Power Consumption total amount and the regional load data that predict the outcome, making data sheet;

And described GPRS communication unit comprises further:

GPRS module: for controlling the actual operation amount of new forms of energy, new forms of energy increase operation amount, then load peak, then load valley newly, these operation of power networks data of Urban Annual Electrical Power Consumption total amount call speed and the amount of calling;

Filtering dividing potential drop anti-jamming circuit: for removing the actual operation amount of new forms of energy, new forms of energy increase operation amount, then load peak, then load valley newly, the noise section of these operation of power networks data of Urban Annual Electrical Power Consumption total amount;

RS485 module: for the actual operation amount of new forms of energy, new forms of energy increase operation amount, then load peak, then load valley newly, these operation of power networks data of Urban Annual Electrical Power Consumption total amount are carried out secondary coding and pass to dual processors unit;

Described filtering dividing potential drop anti-jamming circuit comprises further: by diode D ₁, diode D ₂, diode D ₃, diode D ₄, form anti-interference rectifier circuit; The tie point of diode D1 and diode D2, and be connected in parallel filter inductance L between the tie point of diode D3 and diode D4 ₁, filter capacitor C ₃;

Divider resistance R ₇, divider resistance R ₈with electric capacity of voltage regulation C ₁connect in angle-style, and divider resistance R ₈with electric capacity of voltage regulation C ₁tie point place be connected with the tie point place of diode D2, diode D4; Divider resistance R ₇with electric capacity of voltage regulation C ₁tie point place series diode D ₅with common mode inductance wave filter (ZJYS51); One end of contact resistance C4 between diode D5 and ZJYS51, the other end ground connection of resistance C4; Divider resistance R ₇, divider resistance R ₈tie point place be connected in series divider resistance R9, simultaneously parallel voltage-stabilizing electric capacity C on divider resistance R9 ₂.

The input/output terminal of described ARM embedded microcontroller connects man-machine interaction unit, and its output terminal connects dsp processor;

Other described interactive unit comprises button and LCDs;

Described button connects the first input/output terminal of ARM embedded microcontroller, and described LCDs connects the second input/output terminal of ARM embedded microcontroller.

Present embodiment adopts dual-cpu structure, and dual processors unit comprises ARM embedded microcontroller and dsp processor.Namely by ARM embedded microcontroller 5, as, model is AT91RM9200 and dsp processor 6, if model is that TMS320F28335 forms dual-cpu structure, between two CPU, employing capacity is that the dual port RAM IDT7134 of 4K × 8 carries out exchanges data, and the mode of connection as shown in Figure 2.In the process realized, employ two panels IDT7134, total address space size is 8K × 8, reads and write for DSP and ARM.

GPRS communication unit 2 in present embodiment adopts two-way RS485 circuit 7, and using SP485R chip as transceiver, this chip reliability is very high, and working method is set to semiduplex mode.For preventing signal disturbing to be provided with Phototube Coupling, the chip selected is 6N137, is respectively+3.3V and+5V, as shown in Figure 3,4 at the supply voltage at its two ends.GPIO1 ~ 3 of DSP28335 connect VO, C of RS485 circuit respectively ₁, C ₂three interfaces, 485H with 485L of GPRS-DTU is connected the L of RS485 circuit respectively ₁, L ₂interface.

In order to reduce the interference of external signal to RS485, in external signal before access GPRS-DTU chip 9, also added the filtering dividing potential drop anti-jamming circuit 8 suppressing the ZJYS51 of common mode interference to form, this circuit is by divider resistance R ₇, R ₈, R ₉, electric capacity of voltage regulation C ₁, C ₂with diode D ₁, D ₂, D ₃, D ₄, D ₅the anti-interference rectifier circuit, the filter inductance L that form ₁, filter capacitor C ₃, C ₄form, in this filtering dividing potential drop anti-jamming circuit, diode D ₁, D ₂, D ₃, D ₄form bridge circuit, when there being disturbance, filter inductance L ₁, filter capacitor C ₃automatic filter, diode D ₅with filter capacitor C ₄limited Current flows to, burning voltage; When there is no disturbance, diode D ₁, D ₂, D ₃, D ₄, D ₅, filter inductance L ₁ignore loss and can be considered wire, filter capacitor C ₃, C ₄can be considered short circuit, on circuit without impact; Divider resistance R ₇, R ₈, R ₉composition star connects, divider resistance R ₇, R ₈, electric capacity of voltage regulation C ₁form angle-style to connect, simultaneously ground connection divider resistance R ₉upper parallel voltage-stabilizing electric capacity C ₂, this filtering dividing potential drop anti-jamming circuit effectively can suppress common mode interference and voltage, current signal disturbance, and guarantee the steady operation of GPRS circuit and reduce the distortion rate of wirelessly transmitting data, improve the reliability of data, the baud rate operationally arranged is 9600bps.Power supply designed in present embodiment is that 12V powers, and is realized by power supply unit.

Storage unit 3 in present embodiment, such as, using SD as storage unit, adopts 16G chip external memory, controls, as shown in Figure 5 primarily of CH375 chip.Conveniently the reading of information adds file system herein with write in software program, and data are stored with .TXT file.Direct for storage card grafting PC can process data by staff.SPISCS, GPIO31, SPICLKA, SPISIMOA, SPISOMIA, GPIO32 of DSP28335 are connected to SCS, BZ, SCK, SDI, SDO, INT interface of storage control circuit CH376S chip.

Described man-machine interaction unit adopts VGY12864COLED LCDs 10, and screen size is 128 row × 60 row, can realize image, Charactes Display, and wiring as shown in Figure 6.Operated display unit by button, the interface circuit 11 of button ARM as shown in Figs. 7-8, comprises four operating key upper and lower, left and right, reads historical data, exports data sheet two function keys, and confirms, cancels button.

Adopt above-mentioned containing multi-source electrical network Mid-long term load forecasting device in present embodiment, to the method predicted containing multi-source electrical network Mid-long Term Load, its flow process as shown in Figure 9.Specifically comprise the following steps:

Step 1, by year chooses region factors historical data and network load demand measured value as data sample.

First, present embodiment determines to predict that the time of network load demand is 2013, and user as required, can select the year that will predict voluntarily.

Then, to choose before 2013 the historical data of this regional factor in each time in 6 years and the network load demand measured value corresponding with this year, this area, described region factors historical data comprises following 12 item number certificates: total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price.The actual operation amount of new forms of energy, new forms of energy increase operation amount newly, annual sunshine-duration, annual mean wind speed.Load peak then, then load valley, Urban Annual Electrical Power Consumption total amount.Specifically as shown in Table 1 and Table 2.

As shown in Table 1, the factor historical data affecting this area's electricity consumption for 2005 is as follows: total number of people 201.56 ten thousand, secondary industry total value 86.51 hundred million yuan, gross national product (GNP) 94.71 hundred million yuan, inhabitant's consumption level 4.6 ten thousand yuan/year, average electricity price 0.351 yuan/degree, when the actual operation amount of new forms of energy 8.16 100,011,000, new forms of energy increase operation amount (temporary nothing) newly, annual sunshine-duration 2848.40h/, annual mean wind speed 14.81m/s, load peak 8910.3MVA in 2005, load valley 8247.6MVA then, during Urban Annual Electrical Power Consumption total amount 102.49 hundred million thousand ten thousand.

As shown in Table 2, network load demand measured value in 2005 is 8295.8MVA.

For other time, as 2006 ~ 2013 annual this area factor historical datas can be found shown in 1, again do not repeating.The network load demand measured value corresponding respectively with each time then can be found according to table 2, if 2006 annual network load demand measured values are 8911.5MVA; 2007 annual network load demand measured values are 9564.7MVA, are not again also repeating.

Step 2: pass through to excavate relation between different year, dissimilar data to the sample data of step 1, find the sample weights of region factors sample for workload demand measured value, carry out weight calculation, for determining the influence degree of region factors sample for network load demand measured value of different year.

Described sample weights calculates, and concrete steps are as follows:

(1) build hierarchical structure, described hierarchical structure forms by three layers, builds rule as follows:

Destination layer: be the network load demand measured value that each time before this area's network load to be predicted demand time, at least 5 years is corresponding; Different year is set objectives layer respectively.

In present embodiment, the network load demand time to be predicted using 2013 as this area, when then setting up hierarchical structure, there are 5 different destination layers, be respectively 2008 workload demand measured values as first object layer, within 2009, workload demand measured value is as the second destination layer ..., until workload demand measured value was as the 5th destination layer in 2012.

Rule layer: based on the time of destination layer, in time of 3 years before getting this time, forms the sample weights time.For present embodiment, for first object layer, get before 2008 3 years as first rule layer corresponding with first object layer, namely 2005,2006 and 2007 is the first rule layer.

Each destination layer determines the respective sample weights time, forms respective rule layer respectively:

Second rule layer is: 2006,2007 and 2008;

......

5th rule layer is: 2009,2010 and 2011.

Solution layer: the time determined according to rule layer, calculates the changing sensitivity of each region factors historical data under the change of same load demand measured value under this time; Each solution layer determines different regions factor changing sensitivity separately, forms solution layer respectively.

For present embodiment, for the first rule layer: determine the sensitivity of 12 region factors historical datas in 2005 for workload demand measured value, specifically comprise: total number of people changing sensitivity in 2005, secondary industry total value changing sensitivity in 2005,2005,, Urban Annual Electrical Power Consumption total amount changing sensitivity in 2005.

……

5th rule layer: determine the sensitivity of 12 region factors historical datas in 2009 for workload demand measured value, specifically comprise: total number of people changing sensitivity in 2009, secondary industry total value changing sensitivity in 2009,2009,, Urban Annual Electrical Power Consumption total amount changing sensitivity in 2009.

Set up the complete correlation structure in target time, namely ensure that the arbitrary element in hierarchical structure in every adjacent two layers all has correlativity.

(2) standardization processing is carried out to the data in different levels, is specially:

Destination layer: destination layer represents the problem that will solve, and its data do not need standardization processing.

Rule layer: adopt scale to quantize criterion and carry out standardization processing.Be specially:

The different year workload demand of rule layer does not quantize implication, needs empirically to carry out re-quantization to rule layer.Can according to the difference of workload demand amount between the time be used as quantitative criteria (such as empirically Mid-long Term Load often present and continue the slow situation increased, then can experience think that the relevance of both workload demand amounts is stronger the closer to the time of rule layer).Different year is quantized criterion by 1-9 scale quantize, concrete criterion is: 1-9 scale quantizes criterion and quantizes.1 represents that significance level is the most weak, and 9 represent the strongest.

Solution layer: carry out standardization processing for the every of solution layer respectively according to the determined different year of rule layer, process is consistent with normalized.

Then, 12 sampled data values of n before this time are designated as c _i,j, i=1 ... m, m=12, j=1 ... n, n>=3; Workload demand measured value is designated as a _j, j=1 ... n; Before this time, the quantification in rule layer time is designated as b _j, j=1 ... n; Correlativity between rule layer time and destination layer workload demand value is p _j, solution layer changing sensitivity is designated as q _ij, i=1 ... m, the correlativity between solution layer and rule layer is v _ij, for the time in each rule layer, calculate the every quantized value q in respective party pattern layer by following formula _ij, namely all samples are for the sensitivity of load variations:

$q_{\mathrm{ij}}=\frac{a_j-a_{j-1}}{c_{i,j}-c_{i,j-1}}$

In formula, q _ijimplication is: i-th sample in a jth time is q for the changing sensitivity of a jth time workload demand value _ij; Changing sensitivity reflects, the change of workload demand value and the relation between the adjacent time between sample changed between the adjacent time;

(3) by development of judgment matrix, ask for the weight of each sample of solution layer for the corresponding rule layer time, and rule layer each time is for the weight of destination layer workload demand, is specially:

Have 1 for rule layer judgment matrix A, the time quantity that dimension n is comprised by solution layer determines, wherein element is:

$A=\left[\begin{array}{ccc}\frac{b_1}{b_1}& ...& \frac{b_1}{b_n}\\ ...& & ...\\ \frac{b_n}{b_1}& ...& \frac{b_n}{b_n}\end{array}\right]$

Judgment matrix B for solution layer has n, and its dimension m is determined by the sample size participated in step 1, for a jth matrix wherein element be:

$B_j=\left[\begin{array}{ccc}\frac{q_{1j}}{q_{1j}}& ...& \frac{q_{1j}}{q_{\mathrm{mj}}}\\ ...& & ...\\ \frac{q_{\mathrm{mj}}}{q_{1j}}& ...& \frac{q_{\mathrm{mj}}}{q_{\mathrm{mj}}}\end{array}\right]$

Calculate eigenvalue of maximum and the proper vector of each matrix, tried to achieve the weight vectors of disparity items by proper vector.Weight vectors is designated as respectively:

p＝(p ₁,p ₂,...p _j,...p _n),j＝1..n

v＝(v _1j,v _2j,...,v _ij,...v _mj),i＝1..m

In p expressiveness layer, different year is for the weight of workload demand.Vj represents that different samples in the rule layer jth time are to the weight of workload demand.

(4) calculate the synthetic weights severe of solution layer for destination layer, namely determine the weighing factor of every sample for workload demand value in destination layer time, formula is:

W _i＝Σ _jp _jv _ij

If the training goal in step 3 is prediction regional load requirements in 2013, then sample should comprise sample data and the 2008-2012 workload demand measured value data of 2008-2013.For in 2008-2012 any 1 year sample weights all should with this time for consider, front push away 3 years tissue input data to set up level computational algorithm.Such as 2005-2007 sample data was needed set up sample weights level computational algorithm in 2008 for the sample weights calculating of 2008.Determine the input quantity weights in each time of 2008-2012 successively.

Step 3: adopt the prediction algorithm based on linear mapping, for predicting the annual workload demand total value of somewhere electrical network then, is specially:

(3.1) treat the region factors sample predicting the regional network load demand time, and before this year, the region factors sample in some times, workload demand sample are normalized.

Be normalized the region factors sample in network load demand time to be predicted, in step 1,12 characteristics are sample characteristics, and its time and sample class are in table 1, be specially: total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy increase operation amount newly, annual mean wind speed, annual sunshine-duration, then load peak, load valley then, Urban Annual Electrical Power Consumption total amount.Target is whole society's power load, i.e. workload demand amount.

For the ease of carrying out the normalized of sample quantization coefficient, in quantized samples, choose maximum absolute value value and absolute value minimum value.Such as 2008-2012 total number of people sample, maximal value is 381.69 ten thousand people, and minimum value is 294.10 ten thousand people.

Be normalized by input amendment, normalization formula is:

x＝(x ₀-x _min)/(x _max-x _min)。

In formula, x is the input value after normalized, for affecting 12 feature samples of whole society's total electricity consumption in this example.X ₀for feature samples is without the raw data before normalized, x _maxfor input feature vector sample maximum; x _minfor input feature vector sample minimum.

To the region factors sample evidence different year i after normalized, that determines with step 2 affects weights W _ibe multiplied process.

(3.2) original sample collection is constructed, as the input based on linear mapping prediction algorithm.

With the quantization parameter input value of the characteristic of 12 in table 1 and Analyzing Total Electricity Consumption, according to the various influence factors of bus load, arrange input feature vector data and export Analyzing Total Electricity Consumption data, structure original sample collection.Be input amendment by 12 region factors samples in step (3.1) and full regional load demand is as output quantity set up classified sample set S; Formula is:

$S=\left\{(x_i^0,y_i^0)\right|x_i^0\∈R^L,y_i^0\∈R,i=\mathrm{1,2},...,N\}$

In formula, xi is i-th input vector, represents the sample of 1 year, and its vectorial inside represents the with dividend right data of dissimilar sample; Yi is i-th output valve, represents the workload demand measured value of 1 year; N is that total sample number concentrated by original sample, gets and train for first 5 years in this example.Ordinal number i represents the time of input amendment.

Given training parameter: maximum training iterations T, translational movement δ (uses Weak Classifier to be used as weak learner in present embodiment, the effect of translational movement is by by desired value upper and lower translation one section of fixed range, original sample set is converted to two classified sample sets, thus make learning method become two classification, be convenient to learner and realize weak study, reduce mistake study probability), translational movement is an experience value, is taken as 0.3 in present embodiment.

The classified sample set obtained in step (3.2) is converted to two class classified sample sets, and by described two class sample set sample set combinations, obtain instructing the original sample collection D based on the prediction algorithm of linear mapping, process is as follows:

Recurrence sample set S is converted into classified sample set S1, and method is as follows.

By by desired value y _idistinguish translation fixed range δ up and down, S is converted into two class classified sample sets.

The input vector of classified sample set is made up of 6 impact normalization coefficient input vectors of social power consumption factor and the recurrence sample object value of translation.New desired value is demarcated as 1 and-1 respectively, and corresponding two desired values are the classification samples of 1 and-1, and two class sample sets after conversion are respectively:

$D1=\left\{({(x_i^0,y_i^0+\mathrm{\δ})}^T,+1)\right|x_i^0\∈R^L,y_i^0\∈R,i=\mathrm{1,2},...,N\}$

$D2=\left\{({(x_i^0,y_i^{0-}\mathrm{\δ})}^T,-1)\right|x_i^0\∈R^L,y_i^0\∈R,i=\mathrm{1,2},...,N\}$

In formula, xi and yi implication is identical with in S set, L representative sample number of types, and in this example be 12 kinds, N is the time, is 5 years in this example.

Sample set combination is obtained the input layer of training algorithm, as follows.

D＝{(x _i,y _i)|x _i∈R ^L+1,y _i∈{-1,+1},i＝1,2,...,N}

In formula, x _ifor new samples vector, the xi sample vector in being gathered by original D1 and D2 and the yi after translation form.Y in formula _ifor the desired value of the original sample collection of structure, actual presentation class value is+1 or-1.

If not specified otherwise, sample standard deviation hereinafter represents the sample in sample set D.

(3.3) utilize the sample set D of structure in step (3.2), calculate sample weights vector.Namely initial sample weights vector d is calculated _{1, i}, according to y _ivalue is 1 or-1 sample is considered as two classes, for each sample in each class, carries out weight average distribution according to total sample number amount, and ensures that overall all sample weights sums are 1.

Primary iteration number of times t=1 is set, initializes weights vector d _{1, i}, i represents different sequence number.If y _ithe sample size of=1 is N ₊, y _i=-1 sample size is N _-, d _{1, i}the initial value of inner element arranges as follows.

$d_{1,l,i}=\left\{\begin{array}{c}\frac{1}{2(L+1)N_{+}},(y_i=1)\\ \frac{1}{2(L+1)N_{-}},(y_i=-1)\end{array}\right.,\begin{array}{c}i=\mathrm{1,2}...2N\\ l=\mathrm{1,2}...L+1\end{array}$

D _{1, i}for row vector, wherein l represents different sample types.

By the weight vectors d of each sample _t,iexpected by calculated with weighted average method sample, and obtain sample average m by total weight distribution _t,k.Because input layer sample is all by normalized, therefore sample average should not more than 1.

m _t,k＝{m _t,l,k|l＝1...L+1},k＝1,-1

$m_{t,l,k}=\frac{\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}{x}_{l,i}}{\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}}$

In formula, m _t,kfor dimension and sample x _ibe all the vector of L+1; K=1 or-1, t represent current iteration hierachy number; Corresponding two kinds of representative sample average is categorized as two; Each sample average relies on method of weighted mean to expect to calculate sample by meeting a class sample interior, and obtains divided by the total weight in this classification.

D _{t, l, i}for weight vectors d _t,iin l element; x _l,ifor sample x _iin l element; Add and condition y _i=k represents will meet corresponding y _iall sample elements that value is k sum up according to ordinal number i, hereafter similar.

(3.4) the linear mapping factor is set up.To utilize sample to be 1 be-1 with sample, and sample average constructs sample standard deviation value difference, and constructs between-class scatter and within-class scatter in L+1 dimension space after the conversion, and two dispersions described in utilization, determine the linear mapping factor.

Wherein: between-class scatter S _t,k, formula is as follows:

$S_{t,k}=\frac{\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}{x}_i(x_i-m_{t,k}){(x_i-m_{t,k})}^T}{\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}e}$

Inter _ class relationship characterizes the dispersion degree between different classes of sample, in formula, and d _t,ifor the sample weights of sample in current iteration and iteration layer t is vectorial, its value relies on the right value update in each iteration; x _irepresent sample, e representation dimension and sample x _iidentical vector, its inner element is 1, m _t,kfor the sample average in current iteration; I represents sample sequence number;

Structure within-class scatter S _t,W, formula is as follows:

$S_{t,W}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,i}(x_i-m_{t,k}){(x_i-m_{t,k})}^T$

Within-cluster variance characterizes the dispersion degree between identical category sample.In formula, each element implication is identical with inter _ class relationship formula.

According to linear decision rule, make classification projection mapping x the most accurately _i→ x _i=w _tx _idispersion should be mapped as the equal value difference of sample.Obtain optimum linear mapping-factor best projection maps computing method are the business of inter _ class relationship and within-cluster variance, original sample can be mapped to new sample space, and sample meets inter _ class relationship and the maximization of within-cluster variance ratio within this space.Computing method are:

$w_t^{*}={S_{t,W}}^{-1}(S_{t,1}-S_{t,-1})$

In formula, for best projection to be asked maps, S _t,Wfor within-cluster variance.

(3.5) Weak Classifier is set up.When sample in mapping space is greater than threshold values, be divided into a class, otherwise as another kind of, by the Weak Classifier in best projection map construction current iteration layer, formula is:

h _t＝{h _t,l|l＝1...L+1}

$h_{t,l}\left(x_{i,l}\right)=\left\{\begin{array}{c}+1,{\left(w_t^{*}{x}_i\right)}_l>{\left({\mathrm{\θ}}_t\right)}_l\\ -1,\mathrm{others}\end{array}\right.$

In formula, h _tby being constructed Weak Classifier.X _i,lrepresent x _iin l element, for in l element, l=1 ... L+1 representative sample type.I represents sample sequence number.θ _tbe threshold values vector, determined by the classification error rate in previous iteration layer, (θ _t) _lit is its inner l element.

Weak Classifier is a kind of high efficiency learner, and its form of expression can adjust according to the actual needs.In general the classification value of Weak Classifier is constant, regulates the Rule of judgment of classification can realize different criteria for classifications according to actual needs.

Present embodiment selects a kind of threshold values as criteria for classification, and threshold values is by the classification error rate of a front iteration, and the sample average in adjustment space of linear mapping obtains.The effect of threshold values is for the sorter in current iteration layer provides classification foundation.In present embodiment, the computing formula of threshold value is:

${\mathrm{\θ}}_t=\left\{\begin{array}{c}{w}_t^{*}{m}_t(1-{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}),{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}>{\mathrm{\ϵ}}_{t-1}^{(-1)}\\ w_t^{*}{m}_t(1+{\mathrm{\ϵ}}_{t-1}^{(-1)}),{\mathrm{\ϵ}}_{t-1}^{(-1)}>{\mathrm{\ϵ}}_{t-1}^{\left(1\right)}\end{array}\right.$

In formula, m _tfor mean vector, dimension and sample x _iidentical, inner element is the average of the sample average in current iteration, i.e. m _{t, 1}with m _{t ,-1}average. for meeting y in previous iteration _ithe classification error rate of=1 sample, for meeting y in previous iteration _ithe classification error rate of=-1 sample, error rate of classifying when noting first time iterative computation t=1 is 0.

The Weak Classifier of structure is vector, notes judging that the value of statement is not identical, so the sample be classified all there are differences at every turn at every turn.In general a rational Weak Classifier, due to the existence of linear projection mapping-factor, along with the increase of iterations, classification judges that statement can reduce the mis-classification of Weak Classifier for sample automatically, reduces classification error rate.And then improve the forecasting accuracy of overall prediction algorithm.

h _t＝{h _t,l|l＝1...L+1}

$h_{t,l}\left(x_{i,l}\right)=\left\{\begin{array}{c}+1,{\left(w_t^{*}{x}_i\right)}_l>{\left({\mathrm{\θ}}_t\right)}_l\\ -1,\mathrm{others}\end{array}\right.$

(3.6) by the weight distribution vector of Weak Classifier identification and classification value and desired value contradiction, and by all meet this type of weight distribution add and, obtain classification error rate in current iteration layer, formula is:

${\mathrm{\ϵ}}_t^{\left(1\right)}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}I(h_t{\left(x_i\right)}_l\≠1)$

${\mathrm{\ϵ}}_{t-1}^{(-1)}=\underset{k=1,-1}{\mathrm{\Σ}}\underset{y_i=k}{\mathrm{\Σ}}{d}_{t,l,i}I(h_t{\left(x_i\right)}_l\≠-1)$

In formula, h _t(x _i) represent the classification results of sample, h _t(x _i) _lfor its inside l element, i represents sample sequence number.H _t(x _i) _l≠ 1 expression meets y _ii-th sample x of=1 _iin, l sample elements x _l,iby mis-classification in current iteration layer t, h _t(x _i) _l≠-1 implication is similar.I () represents discriminant function, if the content in bracket is set up, then I=1, otherwise I=0, finally to the weight d of all misjudged samples _{t, l, i}respectively with y _i=1 or y _i=-1 carries out classification weighting, obtains two class classification error rates in current iteration layer.

Often there is mis-classification in Weak Classifier, by calculating classification error rate, can regulate the classifying rules of next time, makes assorting process close to the direction of reducing mis-classification, and this process relies on classification threshold values θ _trealize.

(3.7) ballot parameter alpha is calculated by classification error rate _t, formula is:

${\mathrm{\α}}_t=(1/2)\ln ((1-{\mathrm{\ϵ}}_t)/{\mathrm{\ϵ}}_t),{\mathrm{\ϵ}}_t={\mathrm{\ϵ}}_t^{\left(1\right)}+{\mathrm{\ϵ}}_t^{(-1)}$

In formula, ε _tfor the classification error rate of samples all in current iteration layer.Ballot parameter alpha _tas the t time iteration obtain sorter classification accuracy rate one weigh, when it act as and represents sorter in final weighting all layers, the weight of each layer sorter.In general, the t time Iterative classification error rate is less, represents that the t time Iterative classification device is for overall more effective.

(3.8) upgrade weight vectors by classification error rate and ballot parameter, the weight vectors formula after renewal is as follows:

d _t+1,i＝d _t,i*exp(-α _ty _ih _t(x _i))

In formula, d _t,ifor original weight vectors, α _tfor ballot parameter, y _irepresent score value 1 or-1, its value corresponds to x _imiddle element.The inner element of exp function is vector, and exp take e as the truth of a matter, with each element in vector in bracket for index, carries out exponential function calculating.In formula, No. * represents vector point multiplication, and namely vectorial inner respective items is multiplied, and its result of calculation is the vector that dimension is identical, the d after namely upgrading _{t+1, i}.

Weight in weight vectors should meet and adds and equal 1, therefore is normalized formula to weight vectors and is:

$d_{t+1,i}=\frac{1}{Z_t}{d}_{t+1,i}$

In formula, Zt equal weight vectors interior element adding and.

Repeated execution of steps (3.1) ~ step (3.8), until complete the training of all samples.

(3.9) sorter in final integrated all iteration, obtains regression equation by weighting.I.e. Weak Classifier in the final all iteration layers determined step 3.5, the ballot parameter determined using step 3.7, as respective weights, utilizes the integrated all sorters of weighting, obtains regression equation.This regression equation describes the implicit function relation between sample data and target load requirements, and it is input as 12 master datas in target time, exports the predicted value being area power grid workload demand data;

Regression equation is as follows:

$H\left(x\right)=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{\mathrm{\α}}_t{h}_{t,l}=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{\mathrm{\α}}_t{h}_{t,l}(x^0,y^0)=C$

In formula, y ⁰for comprise the workload demand time to be predicted several years between regional load demand sample vector, x ⁰substantially the region factors sample matrix of the several years sample composition of data is inputted for 12 containing weights coefficient.α _tbe the ballot parameter of the t time iteration, h _t,lrepresent the classification value about l sample in the t time iteration.C is the constant of compressive classification.

H (x) gathers a compressive classification value on D after representing conversion, training result is C, and the unreasonable H (x thought is described ⁰, y ⁰) should more close to C value.Although the value of regression function is the result obtained by sorter h computing, but meets the y of above formula ⁰with x ⁰h (x must be met simultaneously ⁰, y ⁰the equality constraint of)=C regression equation on set D, such as, if wish the prediction workload demand value of 2013, can be considered as parameter by it, forms new for undetermined parameter sample set (x ⁰, y ⁰) and participate in the foundation of regression equation, and make final classification value equal C.

In fact any one can implicit function relation between inner element clear with the integrated regression equation of weighting.Use undetermined coefficient, regional load demand y namely to be predicted ⁰with new region factors sample set x ⁰regression equation is clear, obtain in S set, i.e. load data y ⁰about proper vector x ⁰integrated Models:

y ⁰＝F(x ⁰)

In formula, y ⁰for amount to be predicted, wherein comprise undetermined parameter y ⁰, i.e. workload demand value in 2013; x ⁰it is the sample set (given data group) of 2013 and 12 region factors compositions of front several years.

Substitute into each basic input quantity in target time, after obtaining weights coefficient by step 2, bring in Integrated Models, obtain the predicted value of workload demand data, be the annual gas load total demand of this area in 2013.

Step 4: according to the regional factor of actual area electrical network, utilize the regression equation that step 3 is determined, calculate the workload demand total value of this regional power grid, utilize this composite demand total value, adjust the cooperation of each power generation region in this electrical network, carry out electric network swim scheduling, enable the overall power generation level of this area's electrical network support the overall needs of following regional load.

Although the foregoing describe the specific embodiment of the present invention, the those skilled in the art in this area should be appreciated that these only illustrate, can make various changes or modifications, and do not deviate from principle of the present invention and essence to these embodiments.Scope of the present invention is only defined by the appended claims.

Table 1 is the electricity consumption of 2005-2013 somewhere and correlative factor table thereof.

Table 2 is 2005-2013 somewhere electricity consumption table (workload demand value).

Claims (10)

1., containing a multi-source electrical network Mid-long term load forecasting device, it is characterized in that: in device case, be provided with dual processors unit (1), GPRS communication unit (2), storage unit (3), man-machine interaction unit (4) and power supply unit; Wherein, dual processors unit is connected with GPRS communication unit, man-machine interaction unit, power supply unit, storage unit respectively; And powered by power supply unit.

2. one according to claim 1 is containing multi-source electrical network Mid-long term load forecasting device, it is characterized in that: described man-machine interaction unit (4): for obtaining regional economy class data and weather data, wherein, economic class data comprise: population, secondary industry total value, gross national product (GNP), inhabitant's consumption level and average electricity price; Weather data comprises: annual sunshine-duration, annual mean wind speed;

Described dual processors unit (1): for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors of Urban Annual Electrical Power Consumption total amount carry out affecting intensive analysis, excavate the relation between different year, dissimilar data, the annual workload demand total value of somewhere electrical network is then predicted;

Described GPRS communication unit (2): new forms of energy speculate operation amount, new forms of energy increase operation amount, then load peak, then load valley newly, these service datas of Urban Annual Electrical Power Consumption total amount for obtaining from dispatching center, and input to dual processors unit and calculate; Described dispatching center, i.e. grid dispatching center data server, for prediction unit the actual operation amount of new forms of energy is provided, new forms of energy increase operation amount, then load peak, then load valley newly, these operation of power networks historical datas of Urban Annual Electrical Power Consumption total amount, and obtain this area's load prediction results data;

Described storage unit (3): carrying out summing up editor for total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy being increased newly operation amount, annual sunshine-duration, annual mean wind speed, then load peak, then load valley, these region factors historical datas of Urban Annual Electrical Power Consumption total amount and the regional load data that predict the outcome, making data sheet.

3. one according to claim 1 is containing multi-source electrical network Mid-long term load forecasting device, it is characterized in that: described dual processors unit comprises ARM embedded microcontroller and dsp processor; Described GPRS communication unit (2) adopts two-way RS485 circuit (7), using SP485R chip as transceiver; Also be provided with filtering dividing potential drop anti-jamming circuit in described GPRS communication unit, described filtering dividing potential drop anti-jamming circuit comprises further: by diode D ₁, diode D ₂, diode D ₃, diode D ₄, form anti-interference rectifier circuit; Diode D ₁with diode D ₂tie point, with diode D ₃with diode D ₄tie point between be connected in parallel filter inductance L ₁, filter capacitor C ₃;

Divider resistance R ₇, divider resistance R ₈with electric capacity of voltage regulation C ₁connect in angle-style, and divider resistance R ₈with electric capacity of voltage regulation C ₁tie point place and diode D ₂, diode D4 tie point place be connected;

Divider resistance R ₇with electric capacity of voltage regulation C ₁tie point place series diode D ₅with common mode inductance wave filter ZJYS51;

Diode D ₅and contact resistance C between ZJYS51 ₄one end, resistance C ₄other end ground connection;

Divider resistance R ₇, divider resistance R ₈tie point place be connected in series divider resistance R ₉, simultaneously at divider resistance R ₉upper parallel voltage-stabilizing electric capacity C ₂.

4. one according to claim 1 is containing multi-source electrical network Mid-long term load forecasting device, it is characterized in that: the input/output terminal of described ARM embedded microcontroller connects man-machine interaction unit, and its output terminal connects dsp processor;

Described man-machine interaction unit comprises button and LCDs; Described button connects the first input/output terminal of ARM embedded microcontroller, and described LCDs connects the second input/output terminal of ARM embedded microcontroller.

5., containing a multi-source electrical network Mid-long term load forecasting method, utilize the one described in claim 1 to realize containing multi-source electrical network Mid-long term load forecasting device, it is characterized in that: comprise the steps:

Step 1, by year chooses region factors historical data and network load demand measured value as data sample;

Step 2: the data in integrating step 1, by excavating relation between different year, dissimilar data, find the sample weights of region factors sample for workload demand measured value, for determining the influence degree of the region factors of different year for network load demand;

Step 3: adopt the prediction algorithm based on linear mapping, for predicting the annual workload demand total value of somewhere electrical network then;

Step 4: according to the regional factor of actual area electrical network, utilize the regression equation that step 3 is determined, calculate the workload demand total value of this regional power grid, utilize this composite demand total value, adjust the cooperation of each power generation region in this electrical network, carry out electric network swim scheduling, enable the overall power generation level of this area's electrical network support the overall needs of following regional load.

6. one according to claim 5 is containing multi-source electrical network Mid-long term load forecasting method, it is characterized in that:

Described step 1, by year chooses region factors historical data and network load demand measured value as data sample;

First, the time will predicting network load demand is determined;

Then, after determining this time, the historical data of this regional factor and the network load demand measured value corresponding with this year, this area in each time before being chosen at this time, at least 5 years;

Described region factors historical data comprises following 12 item number certificates: total number of people, secondary industry total value, gross national product (GNP), inhabitant's consumption level, average electricity price, the actual operation amount of new forms of energy, new forms of energy increase operation amount newly, the annual sunshine-duration, annual mean wind speed, load peak then, then load valley, Urban Annual Electrical Power Consumption total amount.

7. one according to claim 5 is containing multi-source electrical network Mid-long term load forecasting method, it is characterized in that:

Described step 2: 12 item number certificates in integrating step 1, by excavating relation between different year, dissimilar data, find the sample weights of region factors sample for workload demand measured value, for determining the influence degree of the region factors of different year for network load demand;

Wherein to the calculating of weight, concrete steps are as follows:

(1) build hierarchical structure, described hierarchical structure forms by three layers, is specially:

Destination layer: be the network load demand measured value that each time before this area's network load to be predicted demand time, at least 5 years is corresponding; Different year is set objectives layer respectively;

Rule layer: based on the time of destination layer, before getting this time, time of at least 3 years, forms the sample weights time; Each destination layer determines the respective sample weights time, forms respective rule layer respectively;

Solution layer: the time determined according to rule layer, calculates the changing sensitivity of each region factors historical data under the change of same load demand measured value under this time; Each solution layer determines different regions factor changing sensitivity separately, forms solution layer respectively;

(2) standardization processing is carried out to the data in different levels, is specially:

Destination layer: destination layer represents the problem that will solve, and its data do not need standardization processing;

Rule layer: adopt scale to quantize criterion and carry out standardization processing;

Solution layer: first, carries out standardization processing according to the determined different year of rule layer, i.e. normalized respectively for the every of solution layer;

Then, the sensitivity of all samples for load variations is determined;

The sensitivity of described load variations reflects, the change of workload demand value and the relation between the adjacent time between sample changed between the adjacent time;

(3) by development of judgment matrix, ask for the weight of each sample of solution layer for the corresponding rule layer time, and rule layer each time is for the weight of destination layer workload demand;

(4) to the synthetic weights severe W of solution layer for destination layer _icalculate, namely determine that every sample in destination layer time affects weights for workload demand value, formula is:

W _i＝Σ _jp _jv _ij

In formula, p _jfor the correlativity between rule layer time and destination layer workload demand value, and there is j=1 ... n, wherein n>=3; v _ijfor the correlativity between solution layer and rule layer, and there is i=1 ... m, wherein, m=12.

8. one according to claim 5 is containing multi-source electrical network Mid-long term load forecasting method, it is characterized in that:

Described step 3: adopt the prediction algorithm based on linear mapping, for predicting the annual workload demand total value of somewhere electrical network then, is specially:

Step 3.1: treat the region factors sample predicting the regional network load demand time, and before this year, the region factors sample in some times, workload demand sample are normalized; And to the region factors sample evidence different year i after normalized, that determines with step 2 affects weights W _ibe multiplied process;

Step 3.2: structure original sample collection, as the input based on linear mapping prediction algorithm;

Be input amendment by 12 region factors samples in step (3.1) and full regional load demand is as output quantity set up classified sample set S; As follows:

In formula, N represents the quantity of input amendment, the representative sample time in sample set S, and L represents initial sample number 12;

The classified sample set obtained in step 3.2 is converted to two class classified sample sets, by described two class sample set combinations, obtains instructing the original sample collection D based on the prediction algorithm of linear mapping; As follows:

D＝{(x _i,y _i)|x _i∈R ^L+1,y _i∈{-1,+1},i＝1,2,...,2N}

In formula, x _irepresent 12 samples and value the formed vector of workload demand measured value after translation in step 1; I represents sample sequence number, and 2N represents the quantity of sample; y _ifor the desired value of the original sample collection of structure, actual value is 1 or-1;

Step 3.3: the sample set D utilizing structure in step 3.2, calculates initial sample weights vector d _{1, i}, according to y _ivalue is 1 or-1 sample is considered as two classes, for each sample in each class, carries out weight average distribution according to total sample number amount, and ensures that overall all sample weights sums are 1;

Set up sample average m _t,k, k=1 or-1, t represent iterations; Corresponding two kinds of representative sample average is categorized as two; Each sample average relies on method of weighted mean to expect to calculate sample by meeting a class sample interior, and obtains divided by the total weight in this classification, and because sample is through normalized, in sample average, any element should not more than 1;

Step 3.4: set up the linear mapping factor;

To utilize sample to be 1 be-1 with sample, and sample average constructs sample standard deviation value difference, and constructs between-class scatter and within-class scatter in L+1 dimension space after the conversion, and two dispersions described in utilization, determine the linear mapping factor; Between-class scatter and within-cluster variance computing formula are:

In formula, S _t,kfor the inter _ class relationship of sample, S _t,Wfor the within-cluster variance of sample; d _t,ifor the sample weights vector of current iteration layer t, its value relies on the right value update in each iteration; E representation dimension and sample x _iidentical vector, its inner element is 1, m _t,kfor two class sample averages in current iteration layer; I represents sample sequence number;

The linear mapping factor computing method are the business of inter _ class relationship and within-cluster variance; Original sample can be mapped to new sample space by the linear mapping factor, and sample meets inter _ class relationship and the maximization of within-cluster variance ratio within this space;

Step 3.5: set up Weak Classifier;

Weak Classifier by best projection map construction current iteration layer: when the sample in mapping space is greater than threshold values, be divided into a class, otherwise as another kind of, the Weak Classifier formula of foundation is:

h _t＝{h _t,l|l＝1...L+1}

In formula, h _tby being constructed Weak Classifier, x _i,lrepresent x _iin l element, for in l element, l=1 ... L+1 representative sample type, i represents sample sequence number, θ _tbe threshold values vector, determined by the classification error rate in previous iteration layer, (θ _t) _lit is its inner l element;

The computing formula of the vector of described threshold value is:

In formula, m _tfor mean vector, dimension and sample x _iidentical, inner element is the average of the sample average in current iteration, i.e. m _{t, 1}with m _{t ,-1}average, for meeting y in previous iteration _ithe classification error rate of=1 sample, for meeting y in previous iteration _ithe classification error rate of=-1 sample, wherein, during first time iterative computation t=1, classification error rate is 0;

Step 3.6: determine classification error rate, formula is:

In formula, h _t(x _i) represent the classification results of sample, h _t(x _i) _lfor its inside l element, i represents sample sequence number, h _t(x _i) _l≠ 1 expression meets y _ii-th sample x of=1 _iin, l sample elements x _l,iby mis-classification in current iteration layer t, h _t(x _i) _l≠-1 implication is similar, and I () represents discriminant function, if the content in bracket is set up, then I=1, otherwise I=0, finally to the weight d of all misjudged samples _{t, l, i}respectively with y _i=1 or y _i=-1 carries out classification weighting, obtains two class classification error rates in current iteration layer;

Step 3.7: utilize the classification error rate in step (3.6) to determine parameter alpha of voting _t, formula is:

In formula, ε _tfor the classification error rate of samples all in current iteration layer; Ballot parameter alpha _tas the t time iteration obtain sorter classification accuracy rate one weigh, when it act as and represents sorter in all layers of final weighting, the weight of each layer sorter, the t time Iterative classification error rate is less, represents that the t time Iterative classification device is for overall more effective;

Step 3.8: upgrade weight vectors by classification error rate and ballot parameter, the weight vectors formula after renewal is as follows:

d _t+1,i＝d _t,i*exp(-α _ty _ih _t(x _i))

In formula, d _t,ifor original weight vectors, y _irepresent score value 1 or-1, its value corresponds to x _imiddle element; The inner element of exp function is vector, and exp take e as the truth of a matter, with each element in vector in bracket for index, carries out exponential function calculating; In formula, No. * represents vector point multiplication, and namely vectorial inner respective items is multiplied, and its result of calculation is the vector that dimension is identical, the d after namely upgrading _{t+1, i};

Weight in weight vectors should meet and adds and equal 1, therefore is normalized formula to weight vectors and is:

In formula, Zt equal weight vectors interior element adding and;

Repeated execution of steps 3.1 ~ step 3.8, until the training completing all samples;

Step 3.9: Weak Classifier in all iteration layers finally step 3.5 determined, the ballot parameter determined using step 3.7 is as respective weights, utilize the integrated all sorters of weighting, obtain regression equation, this regression equation describes the implicit function relation between sample data and target load requirements, it is input as 12 master datas in target time, exports the predicted value being area power grid workload demand data.

9. one according to claim 7 is containing multi-source electrical network Mid-long term load forecasting method, it is characterized in that:

Determine that all samples for the computing method of the sensitivity of load variations are in solution layer in (2) of described step 2:

12 sampled data values of n before this time are designated as c _i,j, i=1 ... m, m=12, j=1 ... n, n>=3; Workload demand measured value is designated as a _j, j=1 ... n; Before this time, the quantification in rule layer time is designated as b _j, j=1 ... n; Correlativity between rule layer time and destination layer workload demand value is p _j, solution layer changing sensitivity is designated as q _ij, i=1 ... m, the correlativity between solution layer and rule layer is v _ij, for the time in each rule layer, calculate the every quantized value q in respective party pattern layer by following formula _ij, namely all samples are for the sensitivity of load variations:

In formula, q _ijimplication is: i-th sample in a jth time is q for the changing sensitivity of a jth time workload demand value _ij.

10. according to claim 7 containing multi-source electrical network Mid-long term load forecasting method, it is characterized in that: described in (3) in described step 2, ask for the weight of each sample of solution layer for the corresponding rule layer time, and rule layer each time is for the weight of destination layer workload demand, be specially:

Have 1 for rule layer judgment matrix A, the time quantity that dimension n is comprised by solution layer determines, wherein element is:

Judgment matrix B for solution layer has n, and dimension m is determined by the sample size participated in step 1, for a jth matrix wherein element be:

Calculate eigenvalue of maximum and the proper vector of each matrix, tried to achieve the weight vectors of disparity items by proper vector, weight vectors is designated as respectively:

p＝(p ₁,p ₂,...p _j,...p _n),j＝1..n

v＝(v _1j,v _2j,...,v _ij,...v _mj),i＝1..m

In formula, in p expressiveness layer, different year is for the weight of workload demand; Vj represents that different samples in the rule layer jth time are to the weight of workload demand.