CN105574607A - Electricity market monthly electricity utilization prediction method - Google Patents
Electricity market monthly electricity utilization prediction method Download PDFInfo
- Publication number
- CN105574607A CN105574607A CN201510917439.5A CN201510917439A CN105574607A CN 105574607 A CN105574607 A CN 105574607A CN 201510917439 A CN201510917439 A CN 201510917439A CN 105574607 A CN105574607 A CN 105574607A
- Authority
- CN
- China
- Prior art keywords
- model
- index
- electricity
- temperature
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000005611 electricity Effects 0.000 title claims abstract description 127
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000007637 random forest analysis Methods 0.000 claims abstract description 27
- 230000000694 effects Effects 0.000 claims abstract description 25
- 238000004364 calculation method Methods 0.000 claims abstract description 12
- 239000011159 matrix material Substances 0.000 claims description 18
- 238000001914 filtration Methods 0.000 claims description 15
- 230000001186 cumulative effect Effects 0.000 claims description 13
- 238000003066 decision tree Methods 0.000 claims description 12
- 230000015572 biosynthetic process Effects 0.000 claims description 11
- 238000003786 synthesis reaction Methods 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 8
- 239000002131 composite material Substances 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000008676 import Effects 0.000 claims description 5
- 230000001276 controlling effect Effects 0.000 claims description 4
- 239000012535 impurity Substances 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000007704 transition Effects 0.000 claims description 4
- 238000001816 cooling Methods 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims description 3
- 238000013524 data verification Methods 0.000 claims description 3
- 238000002474 experimental method Methods 0.000 claims description 3
- 238000010438 heat treatment Methods 0.000 claims description 3
- 239000013589 supplement Substances 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 2
- 230000000295 complement effect Effects 0.000 abstract description 2
- 238000005259 measurement Methods 0.000 abstract 1
- 230000006870 function Effects 0.000 description 23
- 230000000875 corresponding effect Effects 0.000 description 10
- 230000008859 change Effects 0.000 description 7
- PEDCQBHIVMGVHV-UHFFFAOYSA-N Glycerine Chemical compound OCC(O)CO PEDCQBHIVMGVHV-UHFFFAOYSA-N 0.000 description 6
- 230000003111 delayed effect Effects 0.000 description 6
- 229910000831 Steel Inorganic materials 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 4
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 239000010959 steel Substances 0.000 description 4
- 238000007630 basic procedure Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000002045 lasting effect Effects 0.000 description 3
- 230000009021 linear effect Effects 0.000 description 3
- 238000012417 linear regression Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 230000003542 behavioural effect Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000003750 conditioning effect Effects 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000009977 dual effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000003102 growth factor Substances 0.000 description 2
- 239000000047 product Substances 0.000 description 2
- 230000001502 supplementing effect Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241001123248 Arma Species 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004378 air conditioning Methods 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012790 confirmation Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013481 data capture Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009022 nonlinear effect Effects 0.000 description 1
- 230000001373 regressive effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000001932 seasonal effect Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- General Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an electricity market monthly electricity utilization prediction method. The prediction method comprises the following steps of step 1, setting an electricity utilization prediction model, defining a predicted state space model, performing complementary prediction by using a random forest model, loading randomForest, rpart software packages into R software, and leading in an explaining variable and an explained variable; and step 2, determining a prediction model input amount, establishing an air temperature aggregative indicator, adjusting a mobile holiday effect through an effective working day method, predicating a leading indicator, and determining a leading period through model measurement and calculation and coefficient calculation. According to the electricity market monthly electricity utilization prediction method, after the air temperature aggregative indicator, the leading indicator, the business expanding prediction indicator and the like are determined, a state space vector model and a random forest machine study model are combined for performing electricity consumption prediction, so that the prediction method is more accurate and effective.
Description
Technical field
The present invention relates to electricity market power consumption prediction field, be specifically related to a kind of monthly power consumption prediction method of electricity market.
Background technology
One, the main technique methods in existing power consumption prediction and treatment step
1, temperature weighted sum
Due to the otherness of the factors such as physical geography, the temperature that weather bureau announces is issued according to climatic region usually, for there is multiple climatic region, and difference comparatively significantly area, lack temperature record that is unified, reflection general status.In order to make up this defect, we simply add up or weighted sum the temperature of all climatic regions, obtain the overall state of temperature with this, thus obtain overall temperature input quantity for power consumption prediction.If overall temperature is T, the temperature of i-th climatic region is ti, there is the individual different climatic region of n, and so overall temperature can be expressed as:
Wherein, fi is the temperature weight in i-th region.
2, the adjustment of phase in Spring Festival monthly data and merging
The calendar day that the holidays such as lunar calendar Spring Festival occur in calendar month often each year different, we claim this situation to be movable joint holiday.Due to Spring Festival, a large amount of enterprise subtracts stopping production, and power consumption is die-offed, and electricity consumption increase and decrease obviously around the Spring Festival.Therefore, the calendar month occurred along with the Spring Festival is different, and so corresponding monthly electricity consumption data just there will be the variation in each year, and this have impact on precision of prediction greatly.Especially when the Spring Festival is horizontal collapse the Gregorian calendar two months ago after time, adjust electricity consumption just more complicated.Existing disposal route: one, ignores this impact, directly adds up the electricity consumption data of of that month each day as monthly electricity consumption.Namely the actual state of monthly data is directly embodied.Its two, if across two months head and the tail, then merge the data in two month of the place Spring Festival, become the electricity consumption data of month the Spring Festival.In later stage prediction, only use the data of 11 months, predict.Its three, according to the electricity consumption of Spring Festival and the electricity consumption ratio in normal month, as regulation coefficient, back measuring and calculating place month in the Spring Festival, normal power consumption.Circular is:
The of that month daily power consumption of of that month electricity consumption=Σ
Merge moon electricity consumption=Σ place in Spring Festival day moon electricity consumption
Actual electricity consumption/the regulation coefficient of place month in Spring Festival measuring and calculating electricity consumption=place month in the Spring Festival
Regulation coefficient=place month Spring Festival past few years actual electricity consumption/monthly average electricity consumption
3, index system expert assessment method
The existing index to power consumption prediction mainly have chosen macroeconomic variable, industrial trade variable etc.Pass judgment on mainly through experience, expert opinion, the modes such as document tracking determine the power consumption prediction index that will use.
4, time series auto regressive moving average ARIMA model, ARIMA-X model
Time Series AR IMA model carries out based on the self-law of power consumption itself model portrayed.If with electric array { y
tcurrency not only relevant with the past value of self, but also there is certain dependence with it with the external impact e advancing into system, then when portraying this behavioral characteristics with model, both comprised the delayed item of self in model, also comprise external impact in the past, this model is called ARMA model ARIMA (p, d, q), wherein, p is for using electric array { y
tlag order, q is the lag order of external impact e, and d is difference number of times.Its general structure is:
y
t＝φ
1y
t-1+φ
2y
t-2+...+φ
py
t-p+ε
t+θ
1ε
t-1+θ
2ε
t-2+...+θ
qε
t-q
Utilize lag operator, this model can be written as:
Φ(B)y
t＝Θ(B)ε
t
In reality, in order to ensure the stationarity of data, usually first difference is carried out to data, then Modling model, that is: Φ (B) B
dy
t=Θ (B) ε
t, Here it is ARIMA model.
If the currency of sequence is also subject to the impact of its dependent variable, then need to set up ARIMA-X model, that is:
Wherein { y
t, { x
tbe stationary sequence, or through differentiated stationary sequence, { y
tbe power consumption, { x
tfor affecting the factor of electricity consumption.
5, multiple linear regression model
If in order to portray the linear effect of multiple factor to electricity consumption, existing method also utilizes linearity of regression model to predict electricity consumption.Its basic regression expression is:
y
t＝b
0+b
1x
1t+b
2x
2t+...+b
qx
qt+e
t
Wherein q x is the factor affecting electricity consumption, and y is power consumption.By least squares estimate, estimation coefficient b.Then the coefficient of estimation and the value of x is utilized, and the relation that equation is expressed, try to achieve the predicted value of y.
Two, the key step of prior art
1, index set primary election
Passed judgment on by experience, expert opinion, the modes such as document tracking determine the power consumption prediction leading indicators just selected works that will use.Comprising electricity consumption impact external indicator as temperature, economic yield, industrial trade input etc., and reflect endogenous growth factor industry expand index.
2, data processing
Resident's commercial power is very sensitive to temperature, and temperature record itself cannot portray that electricity consumption is non-linear, cumulative bad impact, for the otherness of multiple climatic region, also lacks temperature overall objective.Thus need to adjust into a unified temperature record value as mode input amount.For commercial power monthly data, also to consider the problem of place in Spring Festival month data point reuse.If merge across moon data, then, after needing first the data of the place moon in the Spring Festival to be merged, process as the data of month, then be combined rear monthly data and predict.
3, model specification
If use ARIMA-X model, then first will according to the correlative character before and after the time of electricity sales amount self, the coefficient of autocorrelation of check data sequence and PARCOR coefficients, in order to tentatively to determine the ARIMA model form of matching, i.e. Confirming model auto-correlation lag order p, q.If catch the impact of external factor, need to control in a model the same period or early stage external influence factors.
If linear regression model (LRM), input all main influence factors.What judge in experience tentatively chooses, then by the input variable of the method determination preliminary elections such as related-coefficient test.
4, parameter estimation
Use least square method to carry out the parameter estimation of model, and obtain model residual error estimation item.
5, model evaluation
By the t statistic of estimated parameter, F statistic, fitting coefficient and information criterion etc. determine parameter estimation effect.If ARIMA-X model need by testing model residual error dependency structure, adjustment model seasonal effect in time series dependency structure, until residual error meets normality and non-correlation, obtains preliminary forecast model.
6, model adjustment and prediction
After determining forecast model, available data sample is divided into two parts.The sample of 2/3-3/4 can be used as training sample, and data left can be used as forecast sample.Utilize training sample to predict forecast sample, the actual value of predicted value and forecast sample is compared, carried out the quality of judgment models by predicated error.Finally, the model that error is less becomes final forecast model.
Three, prior art processing mode major defect:
1, on selecting index, existing prediction index is more from outside macroeconomic variable, lacks the index of reflection power industry inherent laws (endogeny).And existing prediction uses same period index to predict more, and namely prediction index and predicted power consumption are the same periods.Current electricity consumption data acquisition is comparatively timely, and other current economic datas compare issue delayed.There is unavailable problem in the data that prediction needs.Especially this prediction based on the index same period, often need first to predict the predictive factors of time span of forecast, re-using predicted data as input variable, predict the electricity consumption of time span of forecast, there is dual prediction in this, and error can expand.
2, data processing have ignored data variation information.First, temperature simply adds up, and have ignored Gradient Effect and the cumulative effect of temperature.Temperature on the impact of electricity consumption only reach more than uniform temperature or under, power load just shows.Such as, about 25 degree air conditioning electricity loads just start to start.In addition, the lasting cumulative bad of high/low temperature weather is also significant on the impact of electricity consumption, is the cumulative bad impact being difficult to portray temperature based on common temperature index and existing model.Non-linear, the lasting cumulative effect of this temperature, needs to reveal this information with the form body of overall temperature just.Secondly, the temperature in Different climate district affects the difference due to the level of economic development, and also there is difference to electricity consumption impact, existing temperature record can not reflect.Therefore, the aggregative index of these impacts of a concentrated expression temperature is needed.Finally, about movable joint holiday, existing calculating, the simple processing mode merged all are difficult to overcome this problem, data fluctuations of having forced joint account level and smooth, have lost information.This fluctuation can affect the precision of month in Spring Festival prediction greatly.Therefore, need to adjust data in the Spring Festival.If can revise in data and index and reflect this change, then the precision of prediction of the applicable data scope of application and model will improve greatly.Even if Use Adjustment coefficient adjusts the method in normal month, also lack portraying Different periods fluctuation after in before the Spring Festival, data point reuse shows slightly coarse.
3, existing model only uses Observable factor, reflects the impact with Electrical change, lacks model catch the variation of unobservable factor.Predictive factors often may not change, and violent variation can occur the variable of the unobservable control of other models, and this have impact on accuracy and the precision of model greatly, and further this influence of change is to next issue certificate.Once the violent variation of unobservable factor becomes the main cause with Electrical change, and model cannot catch, and especially not by this fluctuation deviation of auto modification of given data information, power consumption prediction precision will be made not high.
Summary of the invention
Technical matters to be solved by this invention is to provide a kind of electricity market monthly power consumption prediction method method, structure can reflect the aggregative index that region-wide temperature affects, index needs the notch cuttype impact of portraying electricity consumption, the cumulative effect that high/low temperature is brought, and the Territorial Difference of reflection temperature impact, thus improve the predictive ability of general models.Revise and occur that the data that the fluctuation on date brings are inaccurate in calendar month the Spring Festival, thus catch movable joint Holiday Effect, data core is calculated accurately really.Build the index system with endogenous and leading meaning, the raw factor in electricity consumption that improves portrays ability and power consumption prediction precision.Catch the use electro-mechanical wave that unobservable factor is brought, improve the predictive ability that model changes electricity consumption.
For solving the problems of the technologies described above, the technical solution used in the present invention is as follows:
A kind of electricity market monthly power consumption prediction method method, comprises the following steps:
S1: the setting of power consumption prediction model, comprises state-space model and Random Forest model
S1_11: define the state-space model that will predict
Formula (1) is state transition equation, and formula (2) is observation equation, wherein, and x
tfor unobservable factor, μ
t-qfor external influence factors, t represents the current time, and q represents antephase.Such as, q=1, then t-1 represents with t to be first phase before benchmark, because the delayed needs of external variable calculate confirmation further, therefore represents with q, and the statement of other variant time is similar, such as x
t-1be exactly x
tfront first phase; y
tfor electricity sales amount, Φ is state-transition matrix to be estimated, and Π is observing matrix, and Ψ is the matrix of coefficients of external influence factors, and Γ is the matrix of coefficients that external factor affects sale of electricity, ν
tand ω
tfor Disturbance; And set V and W and be respectively ν
tand ω
tvariance-covariance matrix;
S1_12: by forecast model type in S1_11, determines the parameter Φ needing to estimate, Π, Ψ, Γ, V and W, after namely determining the function coefficient between the variable of forecast model use and variable, load the order bag astsa of R software, wrapped by this order, call corresponding filtering order, i.e. Kalman filtering function Kfilter0, Kfilter1, Kfilter2;
S1_13: import data, data are that electricity sales amount data are as the y in model respectively
t, affect other factor data of electricity sales amount as the μ in model
t, to explanatory variable μ
twith explained variable y
tcarry out standardization;
S1_14: state-space model solve for parameter Φ is set, the initial value of Π, Ψ, Γ, V and W;
S1_15: the likelihood function that state-space model is set, and the parameter Φ of likelihood function, Π, Ψ, Γ, V, W value, uses initial value in S1_14, in likelihood function order, use corresponding filtering order in S1_12, and by filtering order, return likelihood function value;
S1_16: parameter estimation, by parameter Φ, Π, the initial value of Ψ, Γ, V and W, with the likelihood function set as the input item optimizing order optim, use Nelder-Mead method, namely go down the hill Simplex method or Newton-Raphson, i.e. newton-La Fusen alternative manner, the parameter optimal value of iterative computation is as parameter Φ, Π, Ψ, the final value of Γ, V and W;
S1_17: obtain the parameter Φ estimated, after the value of Π, Ψ, Γ, the coefficient as state-space model obtains the final expression formula of forecast model, finally substitutes into variable data and predicts sale of electricity as the new input quantity of model;
S1_21: use Random Forest model to supplement the variable that state-space model is chosen, load randomForest in R software, rpart software package, import explanatory variable and explained variable;
S1_22: setup parameter value N and M: the number N confirming original training sample, determine the variable number M had, sample number is the time interval that Random Forest model estimated parameter needs, and variable number is all factor numbers affecting power consumption;
S1_23: application bootstrap, i.e. Bootstrap sampling method, extract k self-service sample set with putting back to, each self-service sample interval is grown to single decision tree, at each Nodes of tree from a random selecting m variable M variable, from this m feature, choose a feature according to the principle that node impurity level is minimum and carry out dendritic growth;
S1_24: according to k the decision tree classifier generated to the data predicted of needs, namely electricity sales amount data are classified, according to the estimation average of decision tree and the difference of actual average, the little model of selection differences is as final forecast model;
S2: determine forecast model input quantity, i.e. data target, comprises the temperature overall target of reflection electricity consumption staircase effect, cumulative effect, eliminates the monthly data Accounting Problems index of correlation moving effect the Spring Festival and bring, reflect the endogenous index of leading meaning;
S2_11: the structure of temperature overall target, according to June over the years, Dec temperature samming, respectively as high/low temperature average, and fluctuate with samming 2-3 degree and determine temperature threshold range, the high temperature threshold value determined is upt, and low temperature threshold is dpt, when temperature exceeds high temperature threshold value, or lower than low temperature threshold, then cooling or heating load start to start;
S2_12: the Searching point of definite threshold, after setting samming bandwidth, in the scope of samming bandwidth, carries out grid search according to step-length 0.1 degree, afterwards using the measuring and calculating point of each equal warm spot as threshold value;
S2_13: according to threshold value synthesis region temperature index, according to determined threshold value, each high temperature index of the i-th month in area is respectively the temperature on average height of i-th month j day and the difference of high temperature threshold value upt, as input quantity, a value is mapped out again by temperature influence function, add all statistics number of days with i-th month, thus obtain being the high temperature index of i-th month, low temperature index processing mode is identical; Specific formula for calculation is:
wherein, n
ibe the statistics number of days of i-th month, f (*) is temperature influence function;
S2_14: weighting synthesis temperature index, first the temperature in Temperature subarea territory adopts the temperature in the maximum city of this area's power consumption to calculate, secondly, the first two years average that region weight is the resident of place districts and cities, business electricity sales amount accounts for region-wide resident, business electricity sales amount proportion, finally, the weighting synthesis high/low temperature composite index number of region-wide i-th month;
Be specially,
S2_15: determine final region-wide temperature index according to predicated error, all data samples are divided into two parts, wherein 2/3 to 3/4 is training sample, and remaining is experiment sample; According to threshold search point; According to often overlapping the input data of temperature index as state-space model and Random Forest model, these input data are the power consumption data of often overlapping the temperature exponential sum corresponding time; Then according to model prediction program, power consumption is predicted respectively, after prediction, obtain the predicated error of model; Finally according to relative error and the overall average relative error of prediction electricity consumption, consider the temperature index with minimum predicated error, the threshold value that this index is determined is exactly optimum high/low temperature threshold value;
S2_21: by effective method adjustment on working day movable joint Holiday Effect, divide three periods of the Spring Festival, be respectively the first seven sky in the Spring Festival, be i.e. start the first seven sky New Year's Eve; Seven days of Spring Festival, namely the junior one is to the seventh day of lunar month; After the Spring Festival seven days, namely the eighth day of lunar month was to the first month of the lunar year 14; Claim they be before the Spring Festival in rear three periods;
S2_22: the initial regulation coefficient determining Different periods, if the first seven day Spring Festival, regulation coefficient was c1, seven days periods were c2, and latter seven days is c3,
S2_23: the scope of searching for the regulation coefficient of Different periods with step-length 0.1, namely according to the c1 that formula in S2_22 calculates, c2, c3 tri-numerical value are as the initial value of regulation coefficient, numerically search for down at this as step-length according to 0.1, thus the efficiency of formation set, form the coefficient sets of c1, c2 and c3;
S2_24: according to regulation coefficient c1, c2, c3, convert out the number of days of effectively work in three weeks place months after in before the Spring Festival, specifically be calculated as: the last week in the Spring Festival effective working day=number of days to exist i-th month the last week in the Spring Festival of c1*, place week in the Spring Festival effective working day=c2* place in i-th month Spring Festival week exist number of days, after the Spring Festival one week effective working day=i-th month Spring Festival of c3* after the number of days that exists for one week, place month in the Spring Festival effective working day=i-th month non-Spring Festival number of days+the i-th month Spring Festival that work effectively to work number of days;
S2_25: calculate average daily electricity sales amount according to effective set on working day,
S2_26: according to average daily sale of electricity quantity set, as the input data of state-space model and Random Forest model, predicts, according to model predictive error, determines optimal correction coefficient c1 ', c2 ', c3 ';
S2_27: after final regulation coefficient determines effective working day and average daily electricity sales amount, using average daily electricity sales amount as mode input value, the average daily electricity sales amount predicted value obtained, then be multiplied by effective working day, obtain the predicted value of monthly electricity consumption;
S2_31: determine primary election index, affects the leading indicators of electricity sales amount by expert judging method combing, determine antephase by calculation using models and coefficient calculations;
S2_32: influence factor is divided into the factor that will determine antephase and other factors, determine that the factor of antephase is called influence index x, other factors are called Con trolling index z, power consumption is called and is affected index y; First use electricity sales amount y to carry out recurrence to other influences factor z and obtain residual error e
1, then determining that the factor x of antephase returns other influences factor z, obtain residual error e
2, finally calculate e
1and e
2simple correlation coefficient r; Regression equation is:
1)y
t＝Γz
t+e
1
2)x
t＝Πz
t+e
2
Its Chinese style 1) and formula 2) be regression equation, t is time index, and Γ, Π are matrix of coefficients;
S2_33: partial correlation coefficient method is introduced to all indexs, obtains regression residuals e
1and e
2after, according to the computing formula of related coefficient, obtain partial correlation coefficient r:
according to each period of x, calculate the partial correlation coefficient in all periods respectively, select the issue that related coefficient is larger, as the antephase collection of x;
S2_34: all indexs determining antephase are just as the input quantity of forecast model, and according to the External Internal index distinguished, and time dimension constructs laterally longitudinal prediction index system.
According to such scheme, in described step S1_14, initial value is 0, or derives from coefficient regression result or data verification conclusion.
According to such scheme, in step S1_16, maximum iteration time is 1000 times.
According to such scheme, also comprise step S1_18: by the average forecasting error of computational prediction phase, i.e. the mean value of actual value and predicted value difference, judge prediction effect.
According to such scheme, in step S2_11, if time span of forecast is then, then the temperature on average value in June and Dec in winter since calculating the first three years, then on the basis of this samming, each plus-minus 2 degree of bandwidth as threshold value, thus determine the scope of samming threshold value, if samming is T, then samming bandwidth is 2, so the scope of samming threshold value is (T-2, T+2).
According to such scheme, in step S2_13, mapping function chooses out quadratic power,
or absolute value,
or square,
or exponential form,
four kinds of forms.
According to such scheme, in step S2_34, according to step S2_33, determine antephase collection, all antephase set of described index enters model prediction as input quantity, determine predicated error little be final antephase.
Compared with prior art, the invention has the beneficial effects as follows:
The first, the temperature aggregative index of concentrated expression temperature integral status is constructed.First, weather bureau is generally according to the unified status temperature issuing climatic region in the climatic region that physical geography condition is similar.The situation that there is multiple climatic region in administrative area is merely able to announce climatic region temperature separately, and lacks the comprehensive temperature index to multiple climatic region.Secondly, temperature there are differences between zones on electricity consumption impact, and socio-economic development degree is different, and the impact that temperature brings electricity consumption is different.Temperature on electricity consumption impact have Gradient Effect, only in temperature higher or lower than certain threshold value, power load just starts.Tradition utilize temperature directly predict prerequisite be temperature impact be linear change, do not consider this Gradient Effect.Moreover temperature also has cumulative effect to electricity consumption impact, high/low temperature continues cumulative also not identical on electricity consumption impact.Therefore, temperature aggregative index is in structure, we utilize history temperature determination temperature high/low temperature threshold value, sue for peace higher than high temperature threshold value with lower than the dated value of low temperature threshold institute, temperature threshold value and accumulation is utilized to add the General Logistics Department, eventually through the mapping value of function, thus feature gradient and cumulative effect.Finally, we utilize the proportion of region power load in each climatic region electricity consumption as weight, and the weighted calculation temperature of all climatic regions, obtains temperature aggregative index.
The second, effectively working day, method eliminated the data fluctuations problem caused in the variation in Gregorian calendar place month annual holiday in the Spring Festival.The Spring Festival generally appears at the 1-3 month, sometimes crosses over the end of the month and moon head, and this gives the inconvenience adjusted monthly electricity consumption and bring.Electricity consumption in the Spring Festival itself exists regular, but this rule is added on monthly data, adds the variation of calendar day in the Spring Festival, makes monthly data there is fluctuation, affects power consumption prediction.On effective working day, by the regulation coefficient after in before the Spring Festival, adjust monthly effective electricity consumption day, eliminated this influence of fluctuations, thus improved model prediction accuracy.
Three, two shortcomings when leading indicators overcomes the index prediction same period: data can not obtain, conduction meaning is not strong.Use power industry internal indicator, as industry expands index, perfect not enough only by the visual angle of industry exterior angle prediction electric power.First, affect electricity consumption and use the index same period, often when predicting future period, because the data of future period can not obtain, first must predict influence factor, re-using these predicted values and electricity consumption is predicted to thus there is dual forecasting problem.Secondly, use the index same period, often ignore Time Delay and the conduction of factor impact, therefore, current data fluctuations often well can not reflect in next period prediction, and precision of prediction is not high.Finally, existing external indicator mostly is macro-performance indicator, lacks the industry with the endogenous growth factor of power industry and expands index.Therefore, we have selected the industry with the endogenous matter of industry and expand index and have the index of leading meaning, and by partial correlation coefficient method, index is determined the antephase that electricity consumption affects, construct the index system of power consumption prediction, actual prediction improves precision, thus solve power consumption prediction index and there is no endogeny, the shortcoming of leading meaning deficiency.
Four, mutually auxiliary with Random Forest model of state-space model, improves prediction effect.First, state-space model is based on Systems Theory, and the setting of variable interactively agrees with subjective experience, and the model calculation is in order to verify empirical relationship.Random Forest model, completely based on data message itself, does not rely on system and model specification.Both complement each other, and subjectiveness and objectiveness are obtained unified.Secondly, the predictive factors of state-space model is limited, Random Forest model possesses large data analysis capabilities, several factors can be held, especially influence of fluctuations is subject in state-space model predictive factors effect, during precise decreasing, catching more significant influence factor in order to model prediction, is well supplement to state-space model.
Accompanying drawing explanation
Fig. 1 is observed quantity and the quantity of state corresponding relation schematic diagram of state-space model in the present invention.
Fig. 2 is state-space model basic procedure schematic diagram in the present invention.
Fig. 3 is Random Forest model basic procedure schematic diagram in the present invention.
Fig. 4 is the construction step schematic flow sheet of temperature composite index number in the present invention.
Fig. 5 is the schematic flow sheet by effective method adjustment on working day movable joint Holiday Effect in the present invention.
Fig. 6 is leading indicators system computing process schematic in the present invention.
Embodiment
The present invention is power prediction model, is mainly used in prediction electricity needs and electricity market.Its technical problem underlying solved is: one, temperature composite index number, builds the unified composite index number of the region-wide temperature situation of reflection.Two, effectively working day method, solve movable joint Holiday Effect, namely because of inconsistent on the annual Gregorian calendar date in lunar calendar Spring Festival, and the monthly data Accounting Problems caused.Three, leading indicators system, selects the industry of the external indicator of electric power impact and reflection inner factor to expand index based on industrial nature and economic law.By the rational antephase of agriculture products, thus build the external indicator system of power prediction.Four, state-space model and machine learning method combine, on the one hand the strong iterative algorithm of utilization state spatial model, use the large data capture capability of Random Forest model on the other hand, the variable that supplementary state-space model is not considered.
Details are as follows for the specific embodiment of the invention.
(1) forecast model setting: state-space model and Random Forest model
Power sales fluctuation, sometimes from the variation of unobservable factor, needs to catch its behavioral characteristics in time in modeling and forecasting.And this variation is generally difficult to data identification, the effect of this unobservable factor often in data trend sudden change can not be ignored.Therefore, need model to catch this factor on the one hand, what is more important is by the estimation of algorithm realization to this factor, thus the ability of strengthening prediction alteration trend.
State-space model not only controls external influence factors, and carries out dynamic conditioning to the variation of unobservable factor, improves precision of prediction, comparatively accurate to data long-term forecasting.It, by Kalman filtering algorithm, obtains the estimated value of coefficient by data dependence and lastest imformation recurrence.
Supplementing as predictive factors, considers the characteristic of the large data analysis using Random Forest model, catches relevant variable.
First setting Yt is observed quantity, and Xt is quantity of state, and t is current time, and t+1 is the future time instance needing prediction.The basic procedure of state-space model as shown in Figure 1.
To the prediction steps of state-space model be:
1, define the state-space model that will predict, (1) formula is state transition equation, and (2) formula is observation equation.
Wherein, x
tfor unobservable factor, suppose that it is subject to the adaptive impact of delayed first phase, the factor estimated as model exists, not corresponding actual index.μ
tfor external influence factors, in the analysis of reality, be that each affects each prediction index of sale of electricity, select suitable period according to the antephase of each index, time-angle is here designated as t-q, and wherein q is antephase.Y
tfor electricity sales amount, Φ is state-transition matrix to be estimated, it is described that self dynamic conditioning behavior of unobservable factor.Π is observing matrix, it is described that the dynamic effects of unobservable factor to sale of electricity.Ψ is the matrix of coefficients of external influence factors, and show and the contacting of unobservable factor, Γ is the matrix of coefficients that external factor affects sale of electricity, ν
tand ω
tfor Disturbance.
By types of models, determine the parameter needing to estimate, just can be used as parameter required in filtering order below.Parameter in filtering order and model parameter want consistent corresponding, and the model of what type determines the filtering order of what type.
2, after determining the function coefficient between the variable of forecast model use and variable, the order bag astsa of R software is loaded.Wrapped by this order, call relevant Kalman filtering function Kfilter0, Kfilter1, Kfilter2, for subsequent calculations.With the order of 0, be non-time-varying model, and the order used when not having external input quantity, and with 1 order be with sometimes become and external input quantity time the order that uses.Order with 2 is the order that the residual error existed between model uses when being correlated with.
3, import data, standardization is carried out to explanatory variable and explained variable.
4, the initial value of model solve for parameter is set, is set to 0, or other derive from coefficient regression result or data verification conclusion.Such as model variable function coefficient, the element value of transition matrix, and the covariance element value of two equations.These parameters are as the parameter value needed for above-mentioned filter function.
5, the likelihood function of state-space model is set.The parameter value of likelihood function, uses above-mentioned initial parameter value.And, in likelihood function order, use the corresponding filtering order in above-mentioned second step, and by filtering order, return likelihood function value.
6, parameter estimation.By the initial value of parameter, and the likelihood function of setting is as the input item optimizing order optim.Use " Nelder-Mead " or " Newton-Raphson " alternative manner, maximum iteration time is set to 5000 steps.The parameter optimal value of final iterative computation is as the final value of parameter.
7, after obtaining the parameter value that estimates, the coefficient as forecast model obtains the final expression formula of forecast model.Finally substitute into variable data to predict sale of electricity as the new input quantity of model.
8, by the average forecasting error of computational prediction phase, i.e. the mean value of actual value and predicted value difference, judges prediction effect.
Random Forest model is used to carry out supplementing prediction.Random Forest model is based on CART decision Tree algorithms, and each sample needs to carry out downward recurrence classification through decision tree, according to the simple average of the final Output rusults of node impurity level minimum principle.Specific embodiments is as follows:
1, in R software, randomForest is loaded, rpart software package.
2, explanatory variable and explained variable is imported.
3, setup parameter value.Confirm the number N of original training sample, determine the variable number M had.For particular problem, sample number is the time interval that model estimated parameter needs, and variable number is all factor numbers affecting power consumption.Need to determine a definite value m, with deciding when doing decision-making on one node, can use how many variablees, m is less than total variables number M here.Select m=3, mtry=3 is set in order.
4, apply bootsrap bootstrap, extract k self-service sample set with putting back to, in fact, be exactly extracting part time segment on all intervals, build k decision tree, ntree=500 is set.Each self-service sample interval is grown to single decision tree.At each Nodes of tree from a random selecting m variable M variable, from this m feature, choose a feature according to the principle that node impurity level is minimum and carry out dendritic growth.
5, predict and according to k the decision tree classifier generated, the data predicted of needs to be predicted, according to the estimation average of decision tree and the little model of the difference selection differences of actual average as final forecast model.As shown in Figure 2, the idiographic flow that Random Forest model is implemented as shown in Figure 3 for the idiographic flow that state-space model is implemented.
After determining forecast model, need Confirming model input quantity, i.e. data target.Use the temperature overall target of reflection electricity consumption staircase effect, cumulative effect, eliminate the monthly data Accounting Problems moving effect the Spring Festival and bring, build the endogenous index of the leading meaning of reflection simultaneously, the process how realized basic input quantity is described in detail in detail below.
(2) structure of region-wide temperature composite index number
The temperature record of region-wide announcement is divided into several temperature regions, first the temperature of regional is made into temperature index, then weighting synthesizes region-wide aggregative index.
1, temperature threshold range is set
Consider that temperature has steps on electricity consumption impact, and the lasting cumulative bad of high/low temperature impact, the high/low temperature threshold range of a setting temperature, high temperature threshold value is upt, and low temperature threshold is dpt, when temperature exceeds high temperature threshold value, or lower than low temperature threshold, then cooling or heating load start to start.High-temperature load starts general in June, and low temperature stating with load is generally in Dec.
If time span of forecast is then, then the temperature on average value in June and Dec in winter since calculating the first three years.Then on the basis of this samming, each plus-minus 2 degree of bandwidth as threshold value, thus determine the scope of samming threshold value.If samming is T, then samming bandwidth is 2, and so the scope of samming threshold value is (T-2, T+2).Such as, 6-8 month samming is 25 degree, and so 23-27 degree is the scope of high temperature threshold value, the similar process of low temperature threshold.
2, the Searching point of definite threshold
After setting samming bandwidth, in the scope of samming bandwidth, carry out grid search according to step-length 0.1 degree.Such as, between 23-27 degree, there are 40 samming Searching point.Afterwards using each equal warm spot as the step below the measuring and calculating point of threshold value carries out.40 temperature Searching point are determined respectively in such high/low temperature threshold range.
3, according to threshold value synthesis region temperature index
According to determined threshold value, each high temperature index of the i-th month in area is respectively, and the temperature on average height of i-th month j day and the difference of high temperature threshold value upt, as input quantity.Map out a value by temperature influence function again, add all statistics number of days with i-th month, thus obtain being the high temperature index of i-th month.Consider the non-linear effects of temperature, quadratic power is chosen out for mapping function, absolute value, quadratic sum and index four kinds of forms.Low temperature index processing mode is identical, and low temperature threshold is dpt.
Computing formula is:
Wherein, n
ibe the statistics number of days of i-th month, f (*) is temperature influence function, chooses out quadratic power, absolute value, and quadratic sum index four kinds of forms, are specifically calculated as follows:
4, region-wide temperature index is synthesized in weighting
Region-widely be divided into several temperature regions, synthesize region-wide temperature index according to region weight weighting.First the temperature in each temperature region adopts the temperature in the maximum city of this area's electricity sales amount to calculate.Secondly, region weight is the resident of place districts and cities, business electricity sales amount accounts for the first two years average of region-wide resident, business electricity sales amount proportion.Finally, the weighting synthesis high/low temperature composite index number of region-wide i-th month.
5, final region-wide temperature index is determined according to predicated error
All data samples are divided into two parts, and wherein 2/3 to 3/4 is training sample, and remaining is experiment sample.According to threshold search point, have 40 threshold point, the temperature index generated like this has 40 covers.According to often overlapping temperature index, as the input data of state-space model and Random Forest model, then according to model prediction program above, power consumption being predicted respectively, after prediction, obtaining the predicated error of model.Finally according to relative error and the overall average relative error of prediction electricity consumption, consider the temperature index with minimum predicated error, the threshold value that this index is determined is exactly optimum high/low temperature threshold value.Idiographic flow as shown in Figure 4.
(3) effectively working day method, adjustment movable joint Holiday Effect
Around the Spring Festival because plant downtime is returned to work, power consumption is made to possess certain regularity.But there is very large difference again between the date field that the Spring Festival distributes in annual month, thus make to adjust monthly power consumption appearance fluctuation, affect monthly prediction.In order to portray this movable joint Holiday Effect, needing to introduce regulation coefficient and the day electricity of relevant date after in before the Spring Festival is adjusted, to calculate effective working day of electricity consumption.
1, three periods period of the Spring Festival are divided
To go into operation rule according to enterprise, Spring Festival is divided into three periods, is respectively the first seven sky in the Spring Festival (New Year's Eve starts the first seven sky), seven days (junior one) of Spring Festival, after the Spring Festival seven days (the eighth day of lunar month starts).No matter whether the Spring Festival is across the moon, or in any one month, is generally the 1-3 month, then can be described as rear three periods in before the Spring Festival.
2, the initial regulation coefficient of Different periods is determined
Setting regulation coefficient in the first seven day Spring Festival is c1, and Spring Festival is c2 in seven days, and after the Spring Festival seven days is c3.Think the last week of leading portion in the Spring Festival, namely the Spring Festival the last fortnight, power consumption is the electricity consumption of normal condition.According to net power supply day data, since calculating nearly 2 to three years, the average power consumption of the last week in the Spring Festival, and make the last week in Spring Festival electricity obtain a ratio compared with the last fortnight electricity with the Spring Festival, this ratio, just as the adjustment initial coefficients c1 of the first seven day in the Spring Festival, namely represents that the power consumption of the last week in the Spring Festival accounts for the proportion of the last fortnight power consumption in the Spring Festival.Mid-term in the Spring Festival adjusts initial coefficients c2, be then the proportion that the power consumption of Spring Festival accounts for the power consumption of the last fortnight in the Spring Festival.Later stage in Spring Festival adjustment initial coefficients c3 is then the proportion that the power consumption in later stage in the Spring Festival accounts for the power consumption of the last fortnight in the Spring Festival.
3, the scope of the regulation coefficient of Different periods is searched for step-length 0.1
All initial coefficients, search for as before and after step-length according to 0.1, thus the efficiency of formation set.If coefficient value is c1, then add and subtract before and after c1 according to every step-length 0.1, formed newly be worth c1 ± 0.1, and with this value for benchmark continue with 0.1 step iteration.Thus the coefficient sets formed centered by c1.So form the coefficient sets of c1, c2 and c3.
4, according to regulation coefficient set, conversion effective set on working day in the Spring Festival
According to regulation coefficient, convert out the number of days of effectively work in three weeks place months after in before the Spring Festival.Circular is as follows:
The last week in the Spring Festival effective working day=number of days to exist i-th month the last week in the Spring Festival of c1*
Place week in the Spring Festival effective working day=c2* place in i-th month Spring Festival week exist number of days
After the Spring Festival one week effective working day=i-th month Spring Festival of c3* after the number of days that exists for one week
Place month in the Spring Festival effective working day=i-th month non-Spring Festival number of days+the i-th month Spring Festival that work effectively to work number of days
5, average daily electricity sales amount is calculated according to effective set on working day,
6, according to average daily sale of electricity quantity set, as state-space model with the input value entering forest model, predict.According to model predictive error, determine optimal correction coefficient.According to effective working day that final regulation coefficient is corresponding, the average daily electricity sales amount of calculating, as mode input amount, obtains the predicted value of average daily sale of electricity, then is multiplied by effective working day, obtains the predicted value of monthly data.
According to above be the search of 0.1 according to step-length, generate coefficient sets, obtain set on effective working day according to coefficient sets, thus obtain the set of average daily electricity sales amount.Using the input value of these collective datas as forecast model, first matching is carried out to model, by calculating comparison prediction error, Select Error minimum data acquisition determined effective working day is as final effective working days certificate, the regulation coefficient that this working day relies on is just as final regulation coefficient, and idiographic flow as shown in Figure 5.
(4) the leading indicators system predicted
1, primary election index is determined
Affected the leading indicators of electricity sales amount by expert judging method combing, and the industry adding reaction industry inner factor on the basis of external indicator expands index.These indexs rule of thumb judge, and the moving law of industry self has leading meaning, but do not determine whether to possess antephase and antephase issue by practice examining.Therefore, need to determine antephase by calculation using models and coefficient calculations.
2, electricity sales amount and influence factor return other controlling factors respectively
Influence factor is divided into the factor and other factors that will determine antephase.Determine that the factor of antephase is called influence index x, other factors are called Con trolling index z, and power consumption is called and is affected index y.Use electricity sales amount y to carry out recurrence to other influences factor z and obtain residual error e
1, then determining that the factor x of antephase returns other influences factor z, obtain residual error e
2, finally calculate e
1and e
2simple correlation coefficient r.Regression equation is as follows:
1)y
t＝Γz
t+e
1
2)x
t＝Πz
t+e
2
Wherein 1) formula and 2) formula is regression equation, t is time index.Γ, Π are matrix of coefficients.It should be noted that 2) x variable in formula is the influence index will determining antephase, it is the current value of index or front time value etc.Then progressively use it to work as time value to x, delayed first phase, delayed two phases etc. return, and calculate all residual errors.
3, " partial correlation coefficient response method " is introduced to all indexs
Obtain regression residuals e
1and e
2after, according to the computing formula of related coefficient, partial correlation coefficient r can be obtained:
according to each period of x, calculate the partial correlation coefficient in all periods respectively.Select the issue that several related coefficient is larger, as the antephase just selected works of x.
Such as, need to calculate the partial correlation coefficient that steel industry industry has expanded capacity, electricity sales amount is returned other all influence factors such as newly-started area of the commercial house etc., obtains residual error e
1; Again steel industry industry is expanded capacity to return other influences factor, obtain residual error e
2.Finally, above-mentioned formulae discovery e is utilized
1and e
2simple correlation coefficient.This only obtains the partial correlation coefficient that steel industry has expanded this time value and current period electricity sales amount, according to above-mentioned steps, then steel industry has been expanded front first phase, all antephases such as front two phases calculate corresponding partial correlation coefficient respectively, choosing coefficient value larger period, as the antephase just selected works of index.In the antephase collection selected, carry out model prediction respectively, the final little antephase of predicated error of selecting determines final antephase, and idiographic flow as shown in Figure 6.
4, agriculture products system
Determine all indexs of final antephase, divide into outside and internal indicator, and construct laterally longitudinal prediction index system by time dimension.The leading indicators of final establishment is as shown in table 1.
The determined main leading indicators of table 1 power consumption prediction
Claims (7)
1. the monthly power consumption prediction method of electricity market, is characterized in that, comprise the following steps:
S1: the setting of power consumption prediction model, comprises state-space model and Random Forest model
S1_11: define the state-space model that will predict
Formula (1) is state transition equation, and formula (2) is observation equation, wherein, and x
tfor unobservable factor, μ
t-qfor external influence factors, t represents the current time, and q represents antephase, x
t-1x
tfront first phase; y
tfor electricity sales amount, Φ is state-transition matrix to be estimated, and Π is observing matrix, and Ψ is the matrix of coefficients of external influence factors, and Γ is the matrix of coefficients that external factor affects sale of electricity, ν
tand ω
tfor Disturbance; And set V and W and be respectively ν
tand ω
tvariance-covariance matrix;
S1_12: by forecast model type in S1_11, determines the parameter Φ needing to estimate, Π, Ψ, Γ, V and W, after namely determining the function coefficient between the variable of forecast model use and variable, load the order bag astsa of R software, wrapped by this order, call corresponding filtering order, i.e. Kalman filtering function Kfilter0, Kfilter1, Kfilter2;
S1_13: import data, data are that electricity sales amount data are as the y in model respectively
t, affect other factor data of electricity sales amount as the μ in model
t, to explanatory variable μ
twith explained variable y
tcarry out standardization;
S1_14: state-space model solve for parameter Φ is set, the initial value of Π, Ψ, Γ, V and W;
S1_15: the likelihood function that state-space model is set, and the parameter Φ of likelihood function, Π, Ψ, Γ, V and W value, uses initial value in S1_14, in likelihood function order, use corresponding filtering order in S1_12, and by filtering order, return likelihood function value;
S1_16: parameter estimation, by parameter Φ, Π, the initial value of Ψ, Γ, V and W, with the likelihood function set as the input item optimizing order optim, use Nelder-Mead method, namely go down the hill Simplex method or Newton-Raphson, i.e. newton-La Fusen alternative manner, the parameter optimal value of iterative computation is as parameter Φ, Π, Ψ, the final value of Γ, V and W;
S1_17: obtain the parameter Φ estimated, after the value of Π, Ψ, Γ, the coefficient as state-space model obtains the final expression formula of forecast model, finally substitutes into variable data and predicts sale of electricity as the new input quantity of model;
S1_21: use Random Forest model to supplement the variable that state-space model is chosen, load randomForest in R software, rpart software package, import explanatory variable and explained variable;
S1_22: setup parameter value N and M: the number N confirming original training sample, determine the variable number M had, sample number is the time interval that Random Forest model estimated parameter needs, and variable number is all factor numbers affecting power consumption;
S1_23: application bootstrap, i.e. Bootstrap sampling method, extract k self-service sample set with putting back to, each self-service sample interval is grown to single decision tree, at each Nodes of tree from a random selecting m variable M variable, from this m feature, choose a feature according to the principle that node impurity level is minimum and carry out dendritic growth;
S1_24: according to k the decision tree classifier generated to the data predicted of needs, namely electricity sales amount data are classified, according to the estimation average of decision tree and the difference of actual average, the little model of selection differences is as final forecast model;
S2: determine forecast model input quantity, i.e. data target, comprises the temperature overall target of reflection electricity consumption staircase effect, cumulative effect, eliminates the monthly data Accounting Problems index of correlation moving effect the Spring Festival and bring, reflect the endogenous index of leading meaning;
S2_11: the structure of temperature overall target, according to June over the years, Dec temperature samming, respectively as high/low temperature average, and fluctuate with samming 2-3 degree and determine temperature threshold range, the high temperature threshold value determined is upt, and low temperature threshold is dpt, when temperature exceeds high temperature threshold value, or lower than low temperature threshold, then cooling or heating load start to start;
S2_12: the Searching point of definite threshold, after setting samming bandwidth, in the scope of samming bandwidth, carries out grid search according to step-length 0.1 degree, afterwards using the measuring and calculating point of each equal warm spot as threshold value;
S2_13: according to threshold value synthesis region temperature index, according to determined threshold value, each high temperature index of the i-th month in area is respectively the temperature on average height of i-th month j day and the difference of high temperature threshold value upt, as input quantity, a value is mapped out again by temperature influence function, add all statistics number of days with i-th month, thus obtain being the high temperature index of i-th month, low temperature index processing mode is identical; Specific formula for calculation is:
wherein, n
ibe the statistics number of days of i-th month, f (*) is temperature influence function;
S2_14: weighting synthesis temperature index, first subregional temperature adopts the temperature in the maximum city of this area's power consumption to calculate, secondly, the first two years average that region weight is the resident of place districts and cities, business electricity sales amount accounts for region-wide resident, business electricity sales amount proportion, finally, the weighting synthesis high/low temperature composite index number of region-wide i-th month;
Be specially,
S2_15: determine final region-wide temperature index according to predicated error, all data samples are divided into two parts, wherein 2/3 to 3/4 is training sample, and remaining is experiment sample; According to threshold search point; According to often overlapping the input data of temperature index as state-space model and Random Forest model, these input data are the power consumption data of often overlapping the temperature exponential sum corresponding time; Then according to model prediction program, power consumption is predicted respectively, after prediction, obtain the predicated error of model; Finally according to relative error and the overall average relative error of prediction electricity consumption, consider the temperature index with minimum predicated error, the threshold value that this index is determined is exactly optimum high/low temperature threshold value;
S2_21: by effective method adjustment on working day movable joint Holiday Effect, divide three periods of the Spring Festival, be respectively the first seven sky in the Spring Festival, be i.e. start the first seven sky New Year's Eve; Spring Festival seven days, namely the junior one is to the seventh day of lunar month; After the Spring Festival seven days, namely the eighth day of lunar month was to the first month of the lunar year 14, claim they be before the Spring Festival in rear three periods;
S2_22: the initial regulation coefficient determining Different periods, if the first seven day Spring Festival, regulation coefficient was c1, seven days periods were c2, and latter seven days is c3,
S2_23: the scope of searching for the regulation coefficient of Different periods with step-length 0.1, namely according to the c1 that formula in S2_22 calculates, c2, c3 tri-numerical value are as the initial value of regulation coefficient, numerically search for down at this as step-length according to 0.1, thus the efficiency of formation set, form the coefficient sets of c1, c2 and c3;
S2_24: according to regulation coefficient c1, c2, c3, convert out the number of days of effectively work in three weeks place months after in before the Spring Festival, specifically be calculated as: the last week in the Spring Festival effective working day=number of days to exist i-th month the last week in the Spring Festival of c1*, place week in the Spring Festival effective working day=c2* place in i-th month Spring Festival week exist number of days, after the Spring Festival one week effective working day=i-th month Spring Festival of c3* after the number of days that exists for one week, place month in the Spring Festival effective working day=i-th month non-Spring Festival number of days+the i-th month Spring Festival that work effectively to work number of days;
S2_25: average daily electricity sales amount can be calculated according to effective set on working day,
S2_26: according to average daily sale of electricity quantity set, as the input data of state-space model and Random Forest model, predicts, according to model predictive error, determines optimal correction coefficient c1 ', c2 ', c3 ';
S2_27: after final regulation coefficient determines effective working day and average daily electricity sales amount, using average daily electricity sales amount as mode input value, the average daily electricity sales amount predicted value obtained, then be multiplied by effective working day, obtain the predicted value of monthly electricity consumption;
S2_31: determine primary election index, affects the leading indicators of electricity sales amount by expert judging method combing, determine antephase by calculation using models and coefficient calculations;
S2_32: influence factor is divided into the factor that will determine antephase and other factors, determine that the factor of antephase is called influence index x, other factors are called Con trolling index z, power consumption is called and is affected index y; First use electricity sales amount y to carry out recurrence to other influences factor z and obtain residual error e
1, then determining that the factor x of antephase returns other influences factor z, obtain residual error e
2, finally calculate e
1and e
2simple correlation coefficient r; Regression equation is:
1)y
t＝Γz
t+e
1
2)x
t＝Πz
t+e
2
Its Chinese style 1) and formula 2) be regression equation, t is time index, and Γ, Π are matrix of coefficients;
S2_33: partial correlation coefficient method is introduced to all indexs, obtains regression residuals e
1and e
2after, according to the computing formula of related coefficient, obtain partial correlation coefficient r:
according to each period of x, calculate the partial correlation coefficient in all periods respectively, select the issue that related coefficient is larger, as the first selected works of the antephase of x;
S2_34: all indexs determining final antephase, divides into outside and internal indicator, and constructs laterally longitudinal prediction index system by time dimension.
2. the monthly power consumption prediction method of electricity market as claimed in claim 1, it is characterized in that, in described step S1_14, initial value is 0, or derives from coefficient regression result or data verification conclusion.
3. the monthly power consumption prediction method of electricity market as claimed in claim 1, it is characterized in that, in step S1_16, maximum iteration time is 5000 times.
4. the monthly power consumption prediction method of electricity market as claimed in claim 1, is characterized in that, also comprise step S1_18: by the average forecasting error of computational prediction phase, be i.e. the mean value of actual value and predicted value difference, judge prediction effect.
5. the monthly power consumption prediction method of electricity market as claimed in claim 1, it is characterized in that, in step S2_11, if time span of forecast is then, the temperature on average value in June and Dec in winter since then calculating the first three years, then on the basis of this samming, each plus-minus 2 degree of bandwidth as threshold value, thus determine the scope of samming threshold value, if samming is T, then samming bandwidth is 2, and so the scope of samming threshold value is (T-2, T+2).
6. the monthly power consumption prediction method of electricity market as claimed in claim 1, it is characterized in that, in step S2_13, mapping function chooses out quadratic power, namely
or absolute value, namely
or square, namely
or exponential form, namely
four kinds of forms.
7. the monthly power consumption prediction method of electricity market as claimed in claim 1, is characterized in that, according to step S2_33, determine antephase collection, and all antephase collection of described index enters model prediction as input quantity, determine predicated error little be final antephase.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510917439.5A CN105574607A (en) | 2015-12-10 | 2015-12-10 | Electricity market monthly electricity utilization prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510917439.5A CN105574607A (en) | 2015-12-10 | 2015-12-10 | Electricity market monthly electricity utilization prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105574607A true CN105574607A (en) | 2016-05-11 |
Family
ID=55884711
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510917439.5A Pending CN105574607A (en) | 2015-12-10 | 2015-12-10 | Electricity market monthly electricity utilization prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105574607A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107256435A (en) * | 2016-06-30 | 2017-10-17 | 国网江苏省电力公司南通供电公司 | The definite value modification method of predicted value based on platform area daily power consumption |
| CN107273997A (en) * | 2016-06-30 | 2017-10-20 | 国网江苏省电力公司南通供电公司 | The Forecasting Methodology of platform area daily power consumption based on all a few coefficient linear regression model (LRM)s |
| CN107273998A (en) * | 2016-06-30 | 2017-10-20 | 国网江苏省电力公司南通供电公司 | A kind of Temperature correction method predicted for platform area daily power consumption |
| CN107679675A (en) * | 2017-10-20 | 2018-02-09 | 江苏电力信息技术有限公司 | A kind of cash inflow Forecasting Methodology based on moving weighted average |
| CN107862415A (en) * | 2017-11-30 | 2018-03-30 | 成都飞机工业（集团）有限责任公司 | A kind of production prediction method |
| CN110309971A (en) * | 2019-06-28 | 2019-10-08 | 广州市准诚市场咨询有限公司 | A kind of monthly electricity demand forecasting method and system with automatic super rope of engaging in an inspection |
| CN110633860A (en) * | 2019-09-18 | 2019-12-31 | 北京百度网讯科技有限公司 | Maximum demand prediction method, maximum demand prediction device, electronic equipment and storage medium |
| CN111275247A (en) * | 2020-01-14 | 2020-06-12 | 西安理工大学 | Extreme learning machine monthly electric quantity prediction method based on multiple influence factors |
| CN111582568A (en) * | 2020-04-28 | 2020-08-25 | 国网湖南省电力有限公司 | Electric power data-based enterprise rework prediction method during spring festival |
| CN111652424A (en) * | 2020-05-28 | 2020-09-11 | 国网甘肃省电力公司经济技术研究院 | Hydrogen load prediction method |
| CN112258337A (en) * | 2020-09-14 | 2021-01-22 | 陕西讯格信息科技有限公司 | Self-complementing and self-correcting base station energy consumption model prediction method |
| CN112365039A (en) * | 2020-11-03 | 2021-02-12 | 国网冀北电力有限公司计量中心 | Method and system for monthly power consumption prediction based on longitudinal and transverse correlation |
| CN112561226A (en) * | 2020-10-09 | 2021-03-26 | 国网四川省电力公司天府新区供电公司 | Novel high-voltage switch cabinet operation state comprehensive evaluation method |
| CN112734159A (en) * | 2020-12-04 | 2021-04-30 | 国家电网有限公司大数据中心 | Enterprise rework/rework rate calculation method and system |
| CN114707824A (en) * | 2022-03-21 | 2022-07-05 | 国网湖北省电力有限公司电力科学研究院 | Resident refrigerating and heating demand evaluation method based on electric power big data |
| CN117291299A (en) * | 2023-09-22 | 2023-12-26 | 国网江苏省电力有限公司泰州供电分公司 | Moon electricity quantity prediction method considering various influence factors |
| CN118134540A (en) * | 2024-05-07 | 2024-06-04 | 国网浙江省电力有限公司 | Multidimensional data interaction method and system based on energy power supply and demand prediction model |
| CN112734159B (en) * | 2020-12-04 | 2024-09-24 | 国家电网有限公司大数据中心 | Method and system for calculating reworking/reproduction rate of enterprise |
-
2015
- 2015-12-10 CN CN201510917439.5A patent/CN105574607A/en active Pending
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273997B (en) * | 2016-06-30 | 2024-02-02 | 国网江苏省电力公司南通供电公司 | Method and system for predicting daily electricity consumption of platform area based on frequency coefficient linear regression model |
| CN107273997A (en) * | 2016-06-30 | 2017-10-20 | 国网江苏省电力公司南通供电公司 | The Forecasting Methodology of platform area daily power consumption based on all a few coefficient linear regression model (LRM)s |
| CN107273998A (en) * | 2016-06-30 | 2017-10-20 | 国网江苏省电力公司南通供电公司 | A kind of Temperature correction method predicted for platform area daily power consumption |
| CN107256435A (en) * | 2016-06-30 | 2017-10-17 | 国网江苏省电力公司南通供电公司 | The definite value modification method of predicted value based on platform area daily power consumption |
| CN107273998B (en) * | 2016-06-30 | 2024-02-13 | 国网江苏省电力公司南通供电公司 | Air temperature correction method for predicting daily electricity quantity of transformer area |
| CN107256435B (en) * | 2016-06-30 | 2024-02-02 | 国网江苏省电力公司南通供电公司 | Fixed value correction method based on predicted value of daily electricity quantity of station area |
| CN107679675A (en) * | 2017-10-20 | 2018-02-09 | 江苏电力信息技术有限公司 | A kind of cash inflow Forecasting Methodology based on moving weighted average |
| CN107862415A (en) * | 2017-11-30 | 2018-03-30 | 成都飞机工业（集团）有限责任公司 | A kind of production prediction method |
| CN110309971A (en) * | 2019-06-28 | 2019-10-08 | 广州市准诚市场咨询有限公司 | A kind of monthly electricity demand forecasting method and system with automatic super rope of engaging in an inspection |
| CN110633860A (en) * | 2019-09-18 | 2019-12-31 | 北京百度网讯科技有限公司 | Maximum demand prediction method, maximum demand prediction device, electronic equipment and storage medium |
| CN111275247A (en) * | 2020-01-14 | 2020-06-12 | 西安理工大学 | Extreme learning machine monthly electric quantity prediction method based on multiple influence factors |
| CN111275247B (en) * | 2020-01-14 | 2024-02-23 | 西安理工大学 | Extreme learning machine month electric quantity prediction method based on multiple influence factors |
| CN111582568B (en) * | 2020-04-28 | 2023-12-19 | 国网湖南省电力有限公司 | Power data-based method for predicting reworking of enterprises in spring festival |
| CN111582568A (en) * | 2020-04-28 | 2020-08-25 | 国网湖南省电力有限公司 | Electric power data-based enterprise rework prediction method during spring festival |
| CN111652424B (en) * | 2020-05-28 | 2022-06-10 | 国网甘肃省电力公司经济技术研究院 | Hydrogen load prediction method |
| CN111652424A (en) * | 2020-05-28 | 2020-09-11 | 国网甘肃省电力公司经济技术研究院 | Hydrogen load prediction method |
| CN112258337B (en) * | 2020-09-14 | 2024-03-12 | 陕西讯格信息科技有限公司 | Self-complement correction base station energy consumption model prediction method |
| CN112258337A (en) * | 2020-09-14 | 2021-01-22 | 陕西讯格信息科技有限公司 | Self-complementing and self-correcting base station energy consumption model prediction method |
| CN112561226A (en) * | 2020-10-09 | 2021-03-26 | 国网四川省电力公司天府新区供电公司 | Novel high-voltage switch cabinet operation state comprehensive evaluation method |
| CN112365039B (en) * | 2020-11-03 | 2023-11-24 | 国网冀北电力有限公司计量中心 | Method and system for predicting monthly electricity consumption based on longitudinal and transverse correlation |
| CN112365039A (en) * | 2020-11-03 | 2021-02-12 | 国网冀北电力有限公司计量中心 | Method and system for monthly power consumption prediction based on longitudinal and transverse correlation |
| CN112734159A (en) * | 2020-12-04 | 2021-04-30 | 国家电网有限公司大数据中心 | Enterprise rework/rework rate calculation method and system |
| CN112734159B (en) * | 2020-12-04 | 2024-09-24 | 国家电网有限公司大数据中心 | Method and system for calculating reworking/reproduction rate of enterprise |
| CN114707824A (en) * | 2022-03-21 | 2022-07-05 | 国网湖北省电力有限公司电力科学研究院 | Resident refrigerating and heating demand evaluation method based on electric power big data |
| CN114707824B (en) * | 2022-03-21 | 2024-08-13 | 国网湖北省电力有限公司电力科学研究院 | Resident refrigeration and heating demand assessment method based on electric power big data |
| CN117291299A (en) * | 2023-09-22 | 2023-12-26 | 国网江苏省电力有限公司泰州供电分公司 | Moon electricity quantity prediction method considering various influence factors |
| CN117291299B (en) * | 2023-09-22 | 2024-05-03 | 国网江苏省电力有限公司泰州供电分公司 | Moon electricity quantity prediction method considering various influence factors |
| CN118134540A (en) * | 2024-05-07 | 2024-06-04 | 国网浙江省电力有限公司 | Multidimensional data interaction method and system based on energy power supply and demand prediction model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105574607A (en) | Electricity market monthly electricity utilization prediction method | |
| Bang et al. | Fuzzy logic based crop yield prediction using temperature and rainfall parameters predicted through ARMA, SARIMA, and ARMAX models | |
| Crochemore et al. | Bias correcting precipitation forecasts to improve the skill of seasonal streamflow forecasts | |
| Wang et al. | Analysis and application of forecasting models in wind power integration: A review of multi-step-ahead wind speed forecasting models | |
| Fay et al. | On the influence of weather forecast errors in short-term load forecasting models | |
| CN107563554B (en) | Screening method of statistical downscaling model forecasting factors | |
| Antipa et al. | Nowcasting German GDP: A comparison of bridge and factor models | |
| CN106447075B (en) | Industrial electricity demand prediction method and system | |
| CN102495937A (en) | Prediction method based on time sequence | |
| CN109886461A (en) | A kind of Runoff Forecast method and device | |
| CN109657878A (en) | A kind of Air-conditioning Load Prediction method and device | |
| CN109190810B (en) | TDNN-based prediction method for NDVI (normalized difference vegetation index) of northern grassland area of China | |
| CN117745340B (en) | Cigarette market grid capacity rationality prediction method and system based on big data | |
| Littlewood et al. | Predicting daily streamflow using rainfall forecasts, a simple loss module and unit hydrographs: Two Brazilian catchments | |
| JPH04372046A (en) | Method and device for predicting demand amount | |
| CN105389637A (en) | Effective workday construction method for monthly electricity consumption prediction model | |
| CN103308956A (en) | Method for pre-judging future monthly average cloud cover in target area by utilizing cloud climatology data | |
| Ferrara et al. | GDP nowcasting with ragged‐edge data: a semi‐parametric modeling | |
| CN105488595A (en) | Leading index construction method for monthly power consumption prediction model | |
| JP3910948B2 (en) | Information processing method and computer system for power supply plan | |
| Ødegård et al. | Value of information analysis of snow measurements for the scheduling of hydropower production | |
| Wilson et al. | Use of meteorological data for improved estimation of risk in capacity adequacy studies | |
| Lim et al. | Application of expert system to load composition rate estimation algorithm | |
| JPH0922402A (en) | Estimating method for daily maximum power demanded | |
| CN108665090A (en) | Urban distribution network saturation load forecasting method based on principal component analysis Yu Verhulst models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20160511 |