CN104239982A - Method for predicting energy consumption of buildings during holidays and festivals on basis of time series and neural networks - Google Patents

Abstract

The invention provides a method for predicting energy consumption of buildings during holidays and festivals on the basis of time series and neural networks. The method essentially includes predicting energy consumption of the buildings by means of fitting by the aid of the time series; solving prediction errors of energy consumption of the buildings during holidays and festivals; simulating the neural networks by the aid of influence factors on the energy consumption of the buildings during holidays and festivals and the solved prediction errors; computing modification values of the energy consumption of the buildings during holidays and festivals; modifying prediction results of the energy consumption of the buildings during holidays and festivals. The method has the advantages that the energy consumption of the buildings during holidays and festivals can be predicted, and the prediction precision can be improved to a great extent.

Description

A kind of building energy consumption festivals or holidays Forecasting Methodology based on time series and neural network

Technical field

The present invention relates to a kind of Forecasting Methodology of building energy consumption, belong to building energy consumption prediction field, relate to a kind of building energy consumption festivals or holidays Forecasting Methodology based on time series and neural network specifically.

Background technology

Along with China's expanding economy, the problem of office building and large public building highly energy-consuming becomes increasingly conspicuous, and carries out its administration of energy conservation work, has great importance to realization " 12 " building energy conservation object of planning.Building energy conservation is forward position and the study hotspot of current urban construction and social development, prerequisite that comprehensive assessment and analysis is building energy conservation and basis are carried out to the Situation of Heat Consumption of building, and sets up a kind of comparatively simple model that simultaneously can compare Accurate Prediction building energy consumption and seem especially important.

At present, domestic a lot of scholar is studied the Forecasting Methodology of building energy consumption and inquires into.Such as, in document " the buildings electric energy consumption based on BP neural network is predicted ", author adopts LM algorithm, constructs the buildings electricity demand forecasting model based on BP neural network, has carried out Primary Study to Application of Neural Network in the prediction of buildings electric energy consumption.In document " Prediction Model of Urban Building Energy Consumption based on BP Neural Network with Multi-rules & Real-time Training ", author take into account the mechanical periodicity of building electric consumption, add during tectonic network structure month periodic variable and month sequence variables, had further raising to the precision of prediction of building energy consumption.In document " Time Series Method and the application in actual office building energy consumption prediction thereof ", author establishes the building energy consumption forecast model based on Time series analysis method, predicts the energy consumption month by month of certain office building.Various method is by simulating the energy consumption of concrete buildings, predict above, the energy consumption prediction of result display application in normal day has certain feasibility, and predicated error is all in allowed band, but due to building energy consumption influence factor incomplete considered, do not consider the impact of festivals or holidays, make energy consumption prediction result during festivals or holidays not be very desirable.Therefore, the present invention proposes a kind of building energy consumption festivals or holidays Forecasting Methodology based on time series and neural network.

Summary of the invention

For the deficiency of existing structure energy consumption Forecasting Methodology, the present invention proposes the new method of a kind of building energy consumption festivals or holidays based on time series and neural network prediction.Essence of the present invention is, time series is first utilized to carry out matching prediction to building energy consumption, then the predicated error of festivals or holidays is asked for, the energy consumption factor utilizing festivals or holidays and the predicated error solved carry out neuron network simulation, calculate the modified value of energy consumption festivals or holidays, finally the energy consumption prediction result of festivals or holidays is revised, thus realize the energy consumption prediction to building festivals or holidays, and improve its precision of prediction.

The present invention, in order to realize the accurately predicting of building energy consumption festivals or holidays, proposes the new method of a kind of building energy consumption festivals or holidays based on time series and neural network prediction, mainly comprises following steps:

Based on the method for predicting building energy consumption in festivals or holidays situations of time series and neural network, mainly comprise the following steps:

The first step, collects the energy consumption data of this building, and carries out calculus of differences to data, makes it become stable time series Y;

Second step, the stationary sequence Y of the building energy consumption data utilizing the first step to obtain, adopt ARMA model to carry out modeling to it, basic representation is:

(formula 1)

Wherein { Y _t, t=1,2,3 ... N}, Y _t-1..., Y _t-nfor Y is at t-1 ... the value in t-n moment, ε is average is zero, and variance is σ _ε ²stationary white noise, ε _t, ε _t-1..., ε _t-mfor ε is at t, t-1 ... the value in t-m moment, in formula for autoregressive coefficient, θ _j(j=1,2 ..., m) be running mean coefficient;

Utilize this model, according to building energy consumption Stationary Time Series Y t and before the value in moment, the building energy consumption in following t+l moment is made prediction;

3rd step, when calculating festivals or holidays, the predicated error Δ y of building energy consumption, wherein y _t+1for the true power consumption values in t+1 moment, the building energy consumption estimated value in the t+1 moment gone out for utilizing the model prediction in second step;

4th step, set up for predicting when festivals or holidays, the neural network model of building energy consumption predicated error Δ y, comprising the input variable number determining neural network model, the output variable number of neural network model, the hidden layer element number of neural network model, and to neural network model initialization and carry out network training, then, utilize neural network model, under obtaining situation festivals or holidays, the predicated error Δ y of building energy consumption;

5th step, calculate the energy consumption predicted value of building in situation festivals or holidays, its computing formula is: $Y^{*}=\hat{Y}+\mathrm{\Δy}$

Wherein, the energy consumption predicted value obtained for utilizing the model in second step, Δ y is the energy consumption predicated error of festivals or holidays, Y ^*be energy consumption predicted value revised festivals or holidays.

Accompanying drawing explanation

Fig. 1 is the building energy consumption prediction process flow diagram of festivals or holidays in the present invention.

Fig. 2 is the BP neural network structure figure calculating energy consumption modified value festivals or holidays in the present invention.

Embodiment

Below in conjunction with instantiation, be described in detail with reference to the embodiment of Fig. 1 to the inventive method.

Step one: collect data and to go forward side by side line number Data preprocess

This embodiment is a certain office building in Shenzhen, collect the day power consumption data (take seeking time and be at least 2 years) of this building, and holiday information in this period, mean daily temperature, humidity data, and stationary test is carried out to day power consumption time series.If time series is not steady, then by calculus of differences, sample sequence is adjusted, eliminate its tendency and seasonality, make the sequence after changing be stationary sequence.If original day, power consumption time data sequence was { X _t(t=1,2 ..., N), usually, to the cycle be the building of s day power consumption non-stationary, become stable time series { Y after can calculus of differences being carried out _t(t=1,2 ..., N), referred to as Y, wherein calculus of differences is:

Y _t＝X _t-X _t-s

X _tfor data sequence { X _t(t=1,2 ..., N) and in the value of t, X _t-sfor data sequence { X _t(t=1,2 ..., N) and in the value in t-s moment.

Step 2: the steady day power consumption time series { Y that step one is obtained _t(t=1,2 ..., N) and carry out regretional analysis, set up ARMA (n, m) model, specifically comprise:

(1) foundation of model

To the steady day power consumption time series { Y obtained _t(t=1,2 ..., N) and carry out regretional analysis, set up arma modeling:

Wherein { Y _t, t=1,2,3 ... N, Y _t-1..., Y _t-nfor t-1 ... t-n moment stationary sequence Y's

Value, ε is zero-mean, and variance is σ _ε ²stationary white noise, ε _t-1..., ε _t-mfor ε is at t-1 ... the value in t-m moment.In formula for autoregressive coefficient, θ _j(j=1,2 ..., m) be running mean coefficient.

(2) estimation of model parameter

A. the present invention adopts the first estimation of long auto-regression modelling realization to model parameter, and step is as follows:

1. to stationary time series { Y _t(t=1,2 ..., N) and simulate long auto-regression model AR (p),

Z＝φ×W，

Wherein, Z is Stationary Time Series Y value at a time, and W is the value matrix in Stationary Time Series Y other moment before that moment, and φ is its parameter matrix.

Then adopt its parameter matrix of Least Square Method φ, estimated value is denoted as namely for parameter matrix in element, when i is increased to certain numerical value p, all be tending towards 0, then think that the exponent number of autoregression part is p in long auto-regression model.So obtain the inverse function coefficient of long auto-regression model AR (p) with model order p.

2. get the exponent number that n is ARMA (n, m) model AR part, m is the exponent number of MA part.M=1 is made to start search.

3. because of n+m=p, then n=p-m.Solve linear equations, obtains θ _j(j=1,2 ..., m).

$\left[\begin{array}{c}{I}_{n+1}\\ I_{n+2}\\ .\\ .\\ .\\ I_{n+m}\end{array}\right]=\left[\begin{array}{ccccc}{I}_n& I_{n-1}& I_{n-2}& ...& I_{n+1-m}\\ I_{n+1}& I_n& I_{n-1}& ...& I_{n+2-m}\\ ...& ...& ...& ...& ...\\ I_{n+m-1}& I_{n+m-2}& I_{n+m-3}& ...& I_n\end{array}\right]\left[\begin{array}{c}{\mathrm{\θ}}_1\\ {\mathrm{\θ}}_2\\ ...\\ {\mathrm{\θ}}_m\end{array}\right]$

4. check | θ _m| whether be less than 10 ^-6, if not, then make m=m+1, return step and 3. circulate; If be tending towards 0, then the θ once circulated before determining _j(j=1,2 ..., m) be running mean parameter, and the order of MA part is m=m-1, then makes n=p-m perform next step.

5. solve linear equations,

6. check last several value whether be less than 10 ^-6, if so, then save below value, reservation is left for auto-regressive parameter, determine that the order n of AR part equals remaining the number of value, now n<p-m; If not, then need not omit, n=p-m, (i=1,2 ..., n) be auto-regressive parameter.

Just estimated by above-mentioned, obtain for θ in following essence estimation _jwith initial value θ ₁ ⁽⁰⁾, θ ₂ ⁽⁰⁾..., θ _m ⁽⁰⁾

B. in the present invention, the Accuracy extimate of parameter adopts following steps:

1. known stationary time series { Y _t(t=1,2 ..., N), just estimate to obtain model order n through parameter, m and model parameter initial value residual epsilon _tvalue determined by following formula.

2. set k as iterative loop variable, make k=0 start first time iteration.

3. h is estimated by following formula ^(k).

For ARMA (n, m) model, have:

Therefore make y _t=[y _t-1y _t-2y _t-nε _t-1ε _t-2l ε _t-m] ^t, (p=m+n), $f(y_t,\mathrm{\&beta;})=y_t^T\mathrm{\&beta;},v_{\mathrm{it}}^{\left(k\right)}=\frac{\&PartialD;f(y_t,\mathrm{\&beta;})}{\&PartialD;{\mathrm{\&beta;}}_i}{|}_{\mathrm{\&beta;}={\mathrm{\&beta;}}^{\left(k\right)}},(i=\mathrm{1,2},...,p)$

$v^{\left(k\right)}=\left[\begin{array}{cccc}{v}_{1,n+1}^{\left(k\right)}& v_{2,n+1}^{\left(k\right)}& ...& v_{p,n+1}^{\left(k\right)}\\ v_{1,n+2}^{\left(k\right)}& v_{2,n+2}^{\left(k\right)}& ...& v_{p,n+2}^{\left(k\right)}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ v_{1,N}^{\left(k\right)}& v_{2,N}^{\left(k\right)}& ...& v_{p,N}^{\left(k\right)}\end{array}\right]$

wherein β _ifor i-th component of parameter matrix β, i=1,2 ..., p, β ^(k)for the iterative value in parameter matrix β kth generation, k=0,1 ...

4. h is checked ^(k)whether be less than predetermined precision limits δ (p dimensional vector), if h ^(k)< δ, represents that iteration convergence is in claimed range, then with β=β ^(k)+ h ^(k)as the estimated value of model parameter, iterative computation terminates; If h ^(k)< δ is false, then will carry out next iteration calculating.

5. β is made ^(k+1)=β ^(k)+ h ^(k)as the initial value that next iteration calculates, make k=k+1 proceed to step and 3. continue iteration, until h ^(k)< δ iteration terminates.

(3) adaptive test of model

The present invention adopts the Q criterion of model check to test.If the residual error of model of fit is designated as it is ε _testimation.{ ε _tcoefficient of autocorrelation can be calculated by following formula

$\widehat{{\mathrm{\&rho;}}_{\mathrm{\&epsiv;},k}}=\frac{\underset{t=k+1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&epsiv;}}_t{\mathrm{\&epsiv;}}_{t-k}}{\underset{t=k+1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&epsiv;}}_t^2},(k=\mathrm{1,2},...)$

Mathematical statistics can prove, if { ε _t(t=1,2 ... N) be white noise, so statistic then meet the χ that degree of freedom is l-m-n ²distribution.Generally for the sake of simplicity, l-m-n=30 is got, when getting level of confidence and being 95%, by χ ²table checks in χ ²(30) ≈ 44, forms inspection formula Q<=44 like this, when Q value meets this formula, then thinks that corresponding model is applicable models.

(4) data prediction of model

According to (Y _t, Y _t-1...) value to the variable Y in following t+l moment _t+l, (l>0) makes estimation, and estimator is denoted as for ARMA (n, m) model, reverse form according to it ${\mathrm{\&epsiv;}}_t=Y_t-\underset{j=1}{\overset{\&infin;}{\mathrm{\&Sigma;}}}{F}_j{Y}_{t-j},$ Obtaining l step prediction formula is: $\widehat{Y_t}\left(l\right)=\underset{j=1}{\overset{\&infin;}{\mathrm{\&Sigma;}}}{F_j}^{\left(l\right)}{Y}_{t+1-j}$

Wherein, Y _t+1-jfor the whole values of stationary sequence Y before t, coefficient F _j ^(l)can by inverse function { F _j, j=1,2 ... determine, computing formula is as follows:

Inverse function computing formula: θ and be respectively the coefficient in arma modeling

Coefficient F _j ^(l)computing formula: ${F_j}^{\left(l\right)}=\left\{\begin{array}{c}{F}_j,l=1\\ F_{j+l-1}+\underset{i=1}{\overset{l-1}{\mathrm{\&Sigma;}}}{F}_i{F_j}^{(l-i)},l\&NotEqual;1\end{array}\right.$

Step 3: the predicated error Δ y calculating festivals or holidays, wherein y _t+1for the true power consumption values in t+1 moment, for the energy consumption estimated value in t+1 moment.

Step 4: carry out neuron network simulation

Disposal data, determines the input of network, output variable, carries out the normalization of data.Input variable type festivals or holidays divides two kinds, gets 1 and represents day Saturday, gets for 2 representative country's legal festivals and holidays; Temperature gets the medial temperature on same day festivals or holidays, and humidity gets the medial humidity on same day festivals or holidays.Output variable is the energy consumption predicated error Δ y of festivals or holidays.

Then, the relevant information of all increments is formed learning sample, utilizes BP algorithm of neural network, carry out network training according to the network structure of 3 × 4 × 1 shown in Fig. 2.

(1) the input variable number of network is 3, is respectively type festivals or holidays, temperature, humidity.

(2) the output variable number of network is 1, is predicated error Δ y festivals or holidays, namely festivals or holidays energy consumption modified value.

(3) the hidden layer element number of network elects 4 as.

After neural metwork training terminates, just obtain the neural network model calculating energy consumption modified value festivals or holidays, we can according to the relevant information of type festivals or holidays of any one festivals or holidays, temperature and humidity, obtain its energy consumption modified value by this network model, thus carry out correction-compensation to predicting the outcome.

Step 5: the energy consumption predicted value calculating festivals or holidays

The energy consumption predictor calculation formula of festivals or holidays is:

Wherein, for the energy consumption predicted value of above-mentioned arma modeling, Δ y is the energy consumption modified value of the festivals or holidays utilizing neural network to obtain, Y ^*be energy consumption predicted value revised festivals or holidays.

The present invention carries out network analog by utilizing neural network to type festivals or holidays on same day festivals or holidays, temperature, humidity and prediction deviation, obtain the neural network model calculating energy consumption modified value festivals or holidays, thus the energy consumption predicted value of festivals or holidays is revised, improve energy consumption precision of prediction festivals or holidays of buildings to a great extent.

Claims (5)

1., based on the method for predicting building energy consumption in festivals or holidays situations of time series and neural network, mainly comprise the following steps:

The first step, collects the energy consumption data of this building, and carries out calculus of differences to data, makes it become stable time series Y;

Second step, utilizes the stable time series Y of building energy consumption data, and adopt ARMA model ARMA (n, m) to carry out modeling to it, basic representation is:

(formula 1)

Wherein { Y _t, t=1,2,3 ... N}, Y _t-1..., Y _t-nfor Y is at t-1 ... the value in t-n moment, ε is average is zero, and variance is σ _ε ²stationary white noise, ε _t, ε _t-1..., ε _t-mfor ε is at t, t-1 ... the value in t-m moment, in formula for autoregressive coefficient, θ _j(j=1,2 ..., m) be running mean coefficient;

Utilize this model, according to building energy consumption Stationary Time Series Y t and before the value in moment, the building energy consumption in following t+l moment is made prediction;

3rd step, when calculating festivals or holidays, the predicated error △ y of building energy consumption, wherein y _t+1for the true power consumption values in t+1 moment, the building energy consumption estimated value in the t+1 moment gone out for utilizing the model prediction in second step;

4th step, set up for predicting when festivals or holidays, the neural network model of building energy consumption predicated error △ y, comprising the input variable number determining neural network model, the output variable number of neural network model, the hidden layer element number of neural network model, and to neural network model initialization and carry out network training, then, utilize neural network model, under obtaining situation festivals or holidays, the predicated error △ y of building energy consumption;

5th step, calculate the energy consumption predicted value of building in situation festivals or holidays, its computing formula is: wherein, the energy consumption predicted value obtained for utilizing the model in second step, △ y is the energy consumption predicated error of the festivals or holidays utilizing the neural network model in the 4th step to obtain, Y ^*be energy consumption predicted value revised festivals or holidays.

2. the method for claim 1, in second step, adopt long auto-regression modelling to carry out just estimating to the model parameter of formula 1, then, the essence adopting nonlinear least square method to realize the model parameter of formula 1 is estimated.

3. the method for claim 1, wherein in a first step, if the energy consumption data of the building of collecting is { x _t(t=1,2 ..., N), and its cycle be s, then calculus of differences is:

$Y={\&dtri;}_s{X}_t=X_t-X_{t-s},$

Wherein, Y is through differentiated stable time series, for s rank difference operator, X _tfor data sequence is in the value of t, X _t-sfor data sequence is in the value in t-s moment.

4. method as claimed in claim 2, wherein when carrying out just estimating,

Step one, simulates long auto-regression model AR (p) to stationary time series Y,

Z＝φ×W，

Wherein, Z is stationary time series Y value at a time, and W is the value matrix in stationary time series Y other moment before that moment, and φ is its parameter matrix;

Then adopt its parameter matrix of Least Square Method φ, estimated value is denoted as namely for parameter matrix in element, when i is increased to certain numerical value p, all be tending towards 0, then think that the exponent number of autoregression part is p in long auto-regression model, so the model parameter of obtaining (i.e. inverse function) (i=1,2 ..., p) with model order p;

Step 2, gets the exponent number that n is ARMA (n, m) model AR part, and m is the exponent number of MA part; M=1 is made to start search;

Step 3, because of n+m=p, then n=p – m, separates following linear system of equations, obtains θ _j(j=1,2 ..., m):

$\left[\begin{array}{c}{I}_{n+1}\\ I_{n+2}\\ .\\ .\\ .\\ I_{n+m}\end{array}\right]=\left[\begin{array}{ccccc}{I}_n& I_{n-1}& I_{n-2}& ...& I_{n+1-m}\\ I_{n+1}& I_n& I_{n-1}& ...& I_{n+2-m}\\ ...& ...& ...& ...& ...\\ I_{n+m-1}& I_{n+m-2}& I_{n+m-3}& ...& I_n\end{array}\right]\left[\begin{array}{c}{\mathrm{\&theta;}}_1\\ {\mathrm{\&theta;}}_2\\ ...\\ {\mathrm{\&theta;}}_m\end{array}\right]$

Step 4, checks | θ _m| whether level off to 0, if not, then make m=m+1, return step 3 circulation; If be tending towards 0, then the θ once circulated before determining _j(j=1,2 ..., m) be running mean parameter, and the order of MA part is m=m-1, then makes n=p-m perform next step;

Step 5, separates following linear system of equations,

Step 6, checks last several value whether be less than 10 ^-6, if so, then save below value, reservation is left for auto-regressive parameter, determine that the order n of AR part equals remaining the number of value, now n<p-m; If be not less than, then need not omit, now n=p-m, for auto-regressive parameter.

5. the method for claim 1, wherein in second step,

For ARMA (n, m) model, reverse form according to it obtaining l step prediction formula is:

${\hat{Y}}_t\left(l\right)=\underset{j=1}{\overset{\&infin;}{\mathrm{\&Sigma;}}}{F_j}^{\left(l\right)}{Y}_{t+1-j}$

Wherein, Y _t+1-jfor stationary sequence Y is in t+1-j moment value, coefficient F _j ^(l)can by inverse function { F _j, j=1,2 ... determine, computing formula is as follows:

Inverse function computing formula: θ and be respectively the coefficient in arma modeling

Coefficient F _j ^(l)computing formula: ${F_j}^{\left(l\right)}=\left\{\begin{array}{c}{F}_j,l=1\\ F_{j+l-1}+\underset{i=1}{\overset{l-1}{\mathrm{\&Sigma;}}}{F}_i{F_j}^{(l-i)},l\&NotEqual;1\end{array}\right..$