CN104598765A - Building energy consumption prediction method based on elastic adaptive neural network - Google Patents

Abstract

The invention discloses a building energy consumption prediction method based on an elastic adaptive neural network. The method comprises the following steps: selecting historical data of influence factors of building energy consumption, generating input vectors, taking the corresponding historical data of building energy consumption values as output, and obtaining a training sample; training a BP (back propagation) neural network by utilizing the obtained training sample; selecting data of to-be-predicted day of the factors influencing building energy consumption, inputting the data into the BP neural network, and obtaining a prediction value of building energy consumption. The BP neutral network is adopted to realize energy consumption prediction of buildings; an elastic adaptive BP neural network is proposed for a defect that the BP neutral network easily falls into local extremum; the method is used for correcting weight of the BP neutral network, better solves a problem that the BP neutral network easily falls into local extremum, and improves the prediction precision of building energy consumption.

Description

A kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network

Technical field

The present invention relates to a kind of building energy consumption Forecasting Methodology, particularly relate to a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network.

Background technology

Along with the continuous acceleration of urbanization process, energy problem becomes increasingly conspicuous, building energy conservation has become the study hotspot of current social development, comprehensively assessment is carried out to building system energy consumption and comprehensive analysis is prerequisite and the basis of carrying out reducing energy consumption or energy-saving design, and the forecast model setting up the change of reflection energy consumption analyzes understanding building energy consumption Variation and development characteristic for public building energy work from macro-scale to provide effective way and the important means of decision-making foundation.

BP neural network prediction has very strong non-linear mapping capability, be good at finding rule from input and output signal, do not need accurate mathematical model, and computing power is strong, but traditional BP neural network easily causes local minimum, easily there is paralysis and cause learning time long in fixing learning rate in learning process.

For the problems referred to above, the present invention proposes a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network.

Summary of the invention

For solving deficiency of the prior art, the invention provides a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, solve building energy consumption to predict the outcome the low problem of accuracy, and BP neural network is easily absorbed in local extremum, problem that solving precision is low.In order to realize above-mentioned target, the present invention adopts following technical scheme: a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, is characterized in that: comprise step:

(1a) choose the influence factor of building energy consumption, gathering the historical data of building energy consumption influence factor and the building energy consumption historical data corresponding to it, is spring, summer, autumn, winter four groups of training sample sets according to season division;

(1b) input vector is generated according to the historical data of the building energy consumption influence factor in step (1a) respectively by season, using the historical data of corresponding building energy consumption as output data, and input vector and output data are normalized, obtain training sample;

(1c) utilize step (1b) gained training sample to train BP neural network respectively by season, obtain the BP neural network after training;

(1d) gather the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, obtain the prediction input vector after normalized;

(1e) by the BP neural network of the prediction input vector after step (1d) described normalized by input in season correspondence, obtain building energy consumption prediction and export data, prediction exports data obtain day to be predicted building energy consumption predicted value through renormalization process.

Aforesaid a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, is characterized in that: the influence factor of described building energy consumption comprises: the previous day day to be predicted power consumption, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building.

Aforementioned a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, is characterized in that: step (1b) specifically comprises step:

(3a) historical data of gained building energy consumption influence factor is utilized to generate input vector respectively by season, using the historical data of corresponding building energy consumption as output data;

(3b) be normalized respectively step (3a) gained input vector and output data, obtain normalization input vector and normalization output data, wherein the formula of normalized is:

${\stackrel{\‾}{x}}_i=\frac{x_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(1\right)$

$\stackrel{\‾}{y}=\frac{y-y_{\min }}{y_{\max }-y_{\min }}---\left(2\right)$

Wherein, n _ifor input layer number, x _i, before y is respectively normalized, in history input vector, i-th component and history export data, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value, be respectively i-th component and history in the history input vector after normalized and export data;

(3c) the minimum value x of each component in history input vector before normalized is preserved _{i, min}with maximal value x _{i, max}, history exports the minimum value y in data _minwith maximal value y _max.

Aforesaid a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, is characterized in that: the BP neural network structure adopted in described step (1c) comprises input layer, hidden layer and output layer; Described input layer number is 6, comprise power consumption the previous day day to be predicted, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building; Described output layer nodes is 1, is day to be predicted building energy consumption value; Described node in hidden layer m is determined by following formula:

$m=\mathrm{round}(\sqrt{n_{i}l}+0.5)---\left(3\right)$

In formula, n _ifor input layer number, l is output layer nodes, and round () is bracket function.

Aforementioned a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, it is characterized in that: the BP neural network adopted in described step (1c) is a kind of elasticity self-adaptive BP neural networks, its learning rate and weights size adjust by elasticity adaptation rule, and formula is:

$\mathrm{\&eta;}\left(n\right)=\left\{\begin{array}{cc}a\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)<E(n-1)\\ b\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)>c\&times;E(n-1)\end{array}\right.---\left(4\right)$

$w_{\mathrm{ij}}^{(n+1)}=w_{\mathrm{ij}}^{\left(n\right)}+\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}---\left(5\right)$

In formula, η (n) is the learning rate of n-th time, a be greater than 1 constant, b be greater than 0 and be less than 1 constant, c be greater than 1 constant, E (n) is the neuron error of n-th time, is tried to achieve by following formula:

$E\left(n\right)=\frac{1}{2}\underset{k=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_k^2\left(n\right)---\left(6\right)$

Wherein: l is output layer nodes, e _kn () is the error signal of an output layer kth neuron after n iteration, its solution formula is:

e _k(n)＝d _k(n)-m _k(n) (7)

Wherein, d _kn () is the desired output of an output layer kth neuron after n iteration; m _kn () is the actual output of an output layer kth neuron after n iteration;

be the weights of i-th layer of jth neuron (n+1)th time; be the weights variable quantity of i-th layer of jth neuron n-th time, solution formula is:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}-{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ +{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\\ {\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\end{array}\right.---\left(8\right)$

In formula, represent the gradient direction of error surface after n-th iteration; be the weights size updated value of i-th layer of jth neuron n-th time, solution formula is:

${\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}+{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{{\&PartialD;E}^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ 0& \frac{{\&PartialD;E}^{(n+1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\\ -{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{{\&PartialD;E}^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\end{array}\right.---\left(9\right).$

Aforesaid a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, it is characterized in that: described step (1d) gathers the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, normalization formula is:

${\stackrel{\&OverBar;}{x}}_i^{*}=\frac{{x^{*}}_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(10\right)$

Wherein, x ^* _ifor i-th component in the prediction input vector before normalized, for predicting i-th component in input vector after normalized, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized.

Aforesaid a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, it is characterized in that: in step (1e), building energy consumption prediction exports data obtain day to be predicted building energy consumption predicted value through renormalization process, and the formula of renormalization process is:

$y^{*}={\stackrel{\&OverBar;}{y}}^{*}\&times;(y_{\max }-y_{\min })+y_{\min }---\left(11\right)$

for the renormalization obtained through BP neural network prediction building energy consumption predicted data before treatment, y ^*for the building energy consumption predicted value after renormalization process, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value.

The beneficial effect that the present invention reaches: the present invention by the previous day day to be predicted power consumption, buildings location year per capita disposable income, daily forecast solar radiation value to be predicted, daily forecast weather pattern to be predicted, the daily forecast highest temperature to be predicted and building surface volume data, adopt BP neural network, achieve the energy consumption prediction of buildings; The defect of local extremum is easily absorbed in for BP neural network, propose a kind of elasticity adaptation rule improving learning rate and weights, weights for BP neural network correct, and solve the problem that BP neural network is easily absorbed in local extremum preferably, improve the accuracy of system prediction result.

Accompanying drawing explanation

Fig. 1 is the building energy consumption Forecasting Methodology process flow diagram based on elasticity adaptive neural network;

Fig. 2 is BP Artificial Neural Network Structures figure.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.Following examples only for technical scheme of the present invention is clearly described, and can not limit the scope of the invention with this.

As shown in Figure 1, a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network, is characterized in that: comprise step:

(1) choose the influence factor of building energy consumption, gathering the historical data of building energy consumption influence factor and the building energy consumption historical data corresponding to it, is spring, summer, autumn, winter four groups of training sample sets according to season division;

(2) input vector is generated according to the historical data of the building energy consumption influence factor in step (1) respectively by season, using the historical data of corresponding building energy consumption as output data, and input vector and output data are normalized, obtain training sample;

(3) utilize step (2) gained training sample to train BP neural network respectively by season, obtain the BP neural network after training;

(4) gather the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, obtain the prediction input vector after normalized;

(5) by the BP neural network of the prediction input vector after step (4) described normalized by input in season correspondence, obtain building energy consumption prediction and export data, prediction exports data obtain day to be predicted building energy consumption predicted value through renormalization process.

The influence factor of described building energy consumption comprises: the previous day day to be predicted power consumption, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building.When training neural network historical data, certain sky selected is as the Base day (day to be predicted) respectively, by the previous day Base day power consumption, buildings location year per capita disposable income, Base day solar radiation value, Base day weather pattern, the Base day highest temperature and the floor area of building generate input vector, by the Base day (day to be predicted) building energy consumption data as output data, training BP neural network, by practicing the training of a large amount of historical sample, finally obtain the BP neural network after training.

Step (2) specifically comprises step:

1) historical data of gained building energy consumption influence factor is utilized to generate input vector respectively by season, using the historical data of corresponding building energy consumption as output data;

2) to step 1) gained input vector and export data and be normalized respectively, obtain normalization input vector and normalization and export data, wherein the formula of normalized is:

${\stackrel{\&OverBar;}{x}}_i=\frac{x_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(1\right)$

$\stackrel{\&OverBar;}{y}=\frac{y-y_{\min }}{y_{\max }-y_{\min }}---\left(2\right)$

Wherein, x _i, before y is respectively normalized, in history input vector, i-th component and history export data, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value, be respectively i-th component and history in the history input vector after normalized and export data;

3) the minimum value x of each component in history input vector before normalized is preserved _{i, min}with maximal value x _{i, max}, history exports the minimum value y in data _minwith maximal value y _max.

The BP neural network structure adopted in described step (3) comprises input layer, hidden layer and output layer; Described input layer number is 6, comprise power consumption the previous day day to be predicted, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building; Described output layer nodes is 1, is day to be predicted building energy consumption value; Described node in hidden layer is determined by following formula:

$m=\mathrm{round}(\sqrt{n_{i}l}+0.5)---\left(3\right)$

In formula, m is node in hidden layer, n _ifor input layer number, l is output layer nodes, and round () is bracket function.

The BP neural network adopted in described step (3) is a kind of elasticity self-adaptive BP neural networks, and its learning rate and weights size adjust by elasticity adaptation rule, and formula is:

$\mathrm{\&eta;}\left(n\right)=\left\{\begin{array}{cc}a\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)<E(n-1)\\ b\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)>c\&times;E(n-1)\end{array}\right.---\left(4\right)$

$w_{\mathrm{ij}}^{(n+1)}=w_{\mathrm{ij}}^{\left(n\right)}+\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}---\left(5\right)$

In formula, η (n) is the learning rate of n-th time, a be greater than 1 constant, be preferably 1.08; B is greater than 0 constant being less than 1, is preferably 0.65; C be greater than 1 constant, be preferably 1.31; E (n) is the neuron error of n-th time, is tried to achieve by following formula:

$E\left(n\right)=\frac{1}{2}\underset{k=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_k^2\left(n\right)---\left(6\right)$

Wherein: l is output layer nodes, e _kn () is the error signal of an output layer kth neuron after n iteration, its solution formula is:

e _k(n)＝d _k(n)-m _k(n) (7)

Wherein, d _kn () is the desired output of an output layer kth neuron after n iteration; m _kn () is the actual output of an output layer kth neuron after n iteration;

be the weights of i-th layer of jth neuron (n+1)th time; be the weights variable quantity of i-th layer of jth neuron n-th time, solution formula is:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}-{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ +{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\\ {\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\end{array}\right.---\left(8\right)$

In formula, represent the gradient direction of error surface after n-th iteration; be the weights size updated value of i-th layer of jth neuron n-th time, solution formula is:

${\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}+{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{{\&PartialD;E}^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ 0& \frac{{\&PartialD;E}^{(n+1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\\ -{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{{\&PartialD;E}^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{{\&PartialD;E}^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\end{array}\right.---\left(9\right).$

Described step (4) gathers the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, obtain the prediction input vector after normalized, wherein institute's image data comprises: the previous day day to be predicted power consumption, buildings location year per capita disposable income, daily forecast solar radiation value to be predicted, daily forecast weather pattern to be predicted, the daily forecast highest temperature to be predicted and the floor area of building; Described prediction input vector is for constructing vector to institute's image data and being normalized gained normalization input vector, and normalization formula is:

${\stackrel{\&OverBar;}{x}}_i^{*}=\frac{{x^{*}}_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(10\right)$

Wherein, x ^* _ifor i-th component in the prediction input vector before normalized, for predicting i-th component in input vector after normalized, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized.

In described step (5), building energy consumption prediction exports the building energy consumption predicted value obtaining day to be predicted through renormalization process, and the formula of its renormalization process is:

$y^{*}={\stackrel{\&OverBar;}{y}}^{*}\&times;(y_{\max }-y_{\min })+y_{\min }---\left(11\right)$

for the building energy consumption predicted data after the normalized that obtains according to BP neural network, y ^*for the building energy consumption predicted value after renormalization process, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value.

The present invention by the previous day day to be predicted power consumption, buildings location year per capita disposable income, daily forecast solar radiation value to be predicted, daily forecast weather pattern to be predicted, the daily forecast highest temperature to be predicted and the floor area of building, adopt BP neural network, achieve the energy consumption prediction of buildings.The defect of local extremum is easily absorbed in for BP neural network, propose a kind of elasticity adaptation rule improving learning rate and weights, weights for BP neural network correct, and solve the problem that BP neural network is easily absorbed in local extremum preferably, improve the accuracy of system prediction result.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the prerequisite not departing from the technology of the present invention principle; can also make some improvement and distortion, these improve and distortion also should be considered as protection scope of the present invention.

Claims (7)

1., based on a building energy consumption Forecasting Methodology for elasticity adaptive neural network, it is characterized in that: comprise step:

(1a) choose the influence factor of building energy consumption, gathering the historical data of building energy consumption influence factor and the building energy consumption historical data corresponding to it, is spring, summer, autumn, winter four groups of training sample sets according to season division;

(1b) input vector is generated according to the historical data of the building energy consumption influence factor in step (1a) respectively by season, using the historical data of corresponding building energy consumption as output data, and input vector and output data are normalized, obtain training sample;

(1c) utilize step (1b) gained training sample to train BP neural network respectively by season, obtain the BP neural network after training;

(1d) gather the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, obtain the prediction input vector after normalized;

(1e) by the BP neural network of the prediction input vector after step (1d) described normalized by input in season correspondence, obtain building energy consumption prediction and export data, prediction exports data obtain day to be predicted building energy consumption predicted value through renormalization process.

2. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, is characterized in that: the influence factor of described building energy consumption comprises: the previous day day to be predicted power consumption, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building.

3. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, is characterized in that: step (1b) specifically comprises step:

(3a) historical data of gained building energy consumption influence factor is utilized to generate input vector respectively by season, using the historical data of corresponding building energy consumption as output data;

(3b) be normalized respectively step (3a) gained input vector and output data, obtain normalization input vector and normalization output data, wherein the formula of normalized is:

${\stackrel{\&OverBar;}{x}}_i=\frac{x_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(1\right)$

$\stackrel{\&OverBar;}{y}=\frac{y-y_{\min }}{y_{\max }-y_{\min }}---\left(2\right)$

Wherein, n _ifor input layer number, x _i, before y is respectively normalized, in history input vector, i-th component and history export data, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value, be respectively i-th component and history in the history input vector after normalized and export data;

(3c) the minimum value x of each component in history input vector before normalized is preserved _{i, min}with maximal value x _{i, max}, history exports the minimum value y in data _minwith maximal value y _max.

4. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, is characterized in that: the BP neural network structure adopted in described step (1c) comprises input layer, hidden layer and output layer; Described input layer number is 6, comprise power consumption the previous day day to be predicted, buildings location year per capita disposable income, day to be predicted solar radiation value, day to be predicted weather pattern, daily maximum temperature to be predicted and the floor area of building; Described output layer nodes is 1, is day to be predicted building energy consumption value; Described node in hidden layer m is determined by following formula:

$m=\mathrm{round}(\sqrt{n_{i}l}+0.5)---\left(3\right)$

In formula, n _ifor input layer number, l is output layer nodes, and round () is bracket function.

5. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, it is characterized in that: the BP neural network adopted in described step (1c) is a kind of elasticity self-adaptive BP neural networks, its learning rate and weights size adjust by elasticity adaptation rule, and formula is:

$\mathrm{\&eta;}\left(n\right)=\left\{\begin{array}{cc}a\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)<E(n-1)\\ b\&times;\mathrm{\&eta;}(n-1)& E\left(n\right)>c\&times;E(n-1)\end{array}\right.---\left(4\right)$

$w_{\mathrm{ij}}^{(n+1)}=w_{\mathrm{ij}}^{\left(n\right)}+\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}---\left(5\right)$

In formula, η (n) is the learning rate of n-th time, a be greater than 1 constant, b be greater than 0 and be less than 1 constant, c be greater than 1 constant, E (n) is the neuron error of n-th time, is tried to achieve by following formula:

$E\left(n\right)=\frac{1}{2}\underset{k=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_k^2\left(n\right)---\left(6\right)$

Wherein: l is output layer nodes, e _kn () is the error signal of an output layer kth neuron after n iteration, its solution formula is:

e _k(n)＝d _k(n)-m _k(n) (7)

Wherein, d _kn () is the desired output of an output layer kth neuron after n iteration; m _kn () is the actual output of an output layer kth neuron after n iteration;

be the weights of i-th layer of jth neuron (n+1)th time; be the weights variable quantity of i-th layer of jth neuron n-th time, solution formula is:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}-{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ +{\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\\ {\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}& \frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\end{array}\right.---\left(8\right)$

In formula, represent the gradient direction of error surface after n-th iteration; be the weights size updated value of i-th layer of jth neuron n-th time, solution formula is:

${\mathrm{\&Delta;}}_{\mathrm{ij}}^{\left(n\right)}=\left\{\begin{array}{cc}+{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{\&PartialD;E^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}>0\\ 0& \frac{\&PartialD;E^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}=0\\ -{\mathrm{\&Delta;}}_{\mathrm{ij}}^{(n-1)}\&times;\mathrm{\&eta;}\left(n\right)& \frac{\&PartialD;E^{(n-1)}}{\&PartialD;w_{\mathrm{ij}}}\&times;\frac{\&PartialD;E^{\left(n\right)}}{\&PartialD;w_{\mathrm{ij}}}<0\end{array}\right.---\left(9\right).$

6. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, it is characterized in that: described step (1d) gathers the building energy consumption influence factor data genaration prediction input vector of day to be predicted, be normalized, normalization formula is:

${\stackrel{\&OverBar;}{x}}_i^{*}=\frac{{x^{*}}_i-x_{i,\min }}{x_{i,\max }-x_{i,\min }},(i=\mathrm{1,2},...,n_i)---\left(10\right)$

Wherein, x ^* _ifor i-th component in the prediction input vector before normalized, for predicting i-th component in input vector after normalized, x _{i, min}, x _{i, max}be respectively minimum value and the maximal value of i-th component in history input vector before normalized.

7. a kind of building energy consumption Forecasting Methodology based on elasticity adaptive neural network according to claim 1, it is characterized in that: in step (1e), building energy consumption prediction exports data obtain day to be predicted building energy consumption predicted value through renormalization process, and the formula of renormalization process is:

$y^{*}={\stackrel{\&OverBar;}{y}}^{*}\&times;(y_{\max }-y_{\min })+y_{\min }---\left(11\right)$

for the renormalization obtained through BP neural network prediction building energy consumption predicted data before treatment, y ^*for the building energy consumption predicted value after renormalization process, y _min, y _maxbe respectively history before normalized and export minimum value in data and maximal value.