CN106295877A - A kind of intelligent grid electric energy usage amount Forecasting Methodology - Google Patents

Abstract

The open a kind of intelligent grid electric energy usage amount Forecasting Methodology of the present invention, first according to the difference of electricity consumption user types, uses extracting data to be correlated with electric energy usage amount characteristic parameter from electric energy, line number of going forward side by side Data preprocess；Then, building multiple neutral net and predict the electric energy usage amount of each electricity consumption type of service respectively, the parameter of each neutral net is generated by a kind of fast algorithm；Finally, total electric energy usage amount of prediction it is calculated by each electricity consumption type of service electric energy usage amount, the method of the present invention takes into full account that electric energy usage amount is comprised dissimilar electricity consumption behavior and the inherent space time correlation between them, solve the traditional prediction method restricted problem to precision of prediction, and the method for the present invention is capable of predicting quickly and accurately.

Description

A kind of intelligent grid electric energy usage amount Forecasting Methodology

Technical field

The present invention relates to electric power network field, particularly to a kind of intelligent grid electric energy usage amount Predicting Technique.

Background technology

Ensure that the equilibrium of supply and demand that electric energy uses is a major issue in electrical network.When supply electric energy is less than power demand Time, it is likely to result in the interruption of some electrical power service, even causes large-scale power failure；If supply electric energy is more than power demand Time, this unnecessary part electric energy needs to transmit or use the energy storage device of suitable size to store extraly, transmits, stores up Can and their maintenances of causing etc. can increase the cost that electric energy uses significantly.For tackling this problem, it is necessary for electrical network Electric energy usage amount predict quickly and accurately, thus ensure the equilibrium of supply and demand that electric energy uses, improve the economy of electric power networks Benefit and social benefit.

Electric energy in electrical network uses and mainly includes commercial power, farming power, residential electricity consumption and commercial power.Wherein, work Industry electricity consumption, farming power typically have clear and definite plan due to its production process, and its electric energy uses has the strongest regularity.Its Its electricity consumption, such as residential electricity consumption, commercial power etc., owing to user types is various, number of users dynamically changes, electricity consumption is planned poor Etc. reason, its electric energy usage amount rule is inconspicuous, it is difficult to the most predicted.

Intelligent grid is the electric power networks of a kind of modernization.Relative to tradition electrical network, intelligent grid is relatively reliable, safe, Efficiently, and it may use that the sensing of advanced person, measure, the technology such as communication obtains the most careful electric energy and uses data, thus helps Help electrical network that electric energy usage amount is more accurately predicted.

At present, in electrical network, existing electric energy usage amount Forecasting Methodology is primarily present the following aspects problem:

1, precision of prediction is the highest.In electrical network, electric energy usage amount changes over the time series of composition by economic development, product The multifactor impacts such as industry structure, weather, and single electricity consumption behavior has uncertainty so that have between electricity consumption behavior Complicated non-linear relation；Traditional time series analysis and Predicting Technique are difficult to accurately reflect power consumption and change contained electricity consumption Complex nonlinear relation between behavior, thus affect the precision of prediction.

2, adaptability or real-time are poor.Existing electric energy usage amount Forecasting Methodology mostly use mathematical model to predict into Row modeling, modeling process is more complicated, and when some factor affecting electric energy usage amount changes, relevant parameter can not From main modulation；Some Forecasting Methodology has preferably adaptivity, but training/calculation cost is higher, and real-time is poor.

The most common, to can be used for the prediction of electric energy usage amount method mainly includes predicting based on seasonal effect in time series Method and Forecasting Methodology based on regression analysis.

1. based on seasonal effect in time series Forecasting Methodology: by finding historical data certain rule time dependent, set up and close Suitable mathematical model is to be predicted.As based on autoregression model (Auto-Regressive model is called for short AR model) pre- Survey method, Forecasting Methodology based on moving average model (Moving-Average model, MA model) model, based on autoregression The Forecasting Methodology etc. of moving average model (Auto-Regressive Moving-Average model, arma modeling).

2. Forecasting Methodology based on regression analysis: by historical data is carried out regression analysis, set up electric energy usage amount with One group model equation of other correlated variables is to be predicted.As based on support vector machine (Support Vector Machine, SVM) Forecasting Methodology, Forecasting Methodology etc. based on k nearest neighbor (K-Nearest Neighbor, K-NN).

Above method is the most only come electric energy by the history value of electric energy usage amount and the relation of predictive value, historical variations rule Usage amount is predicted, and does not considers the comprised electricity consumption behavior of electric energy usage amount and the internal relation between them (e.g., Relation between electric energy usage amount and the electric energy usage amount of computer of electric light in company), thus limit the precision of prediction.

Summary of the invention

The present invention solves above-mentioned technical problem, it is proposed that a kind of intelligent grid electric energy usage amount Forecasting Methodology, first root According to the difference of electricity consumption user types, use extracting data to be correlated with electric energy usage amount characteristic parameter from electric energy, and it is pre-to carry out data Process；Then, build multiple neutral net and predict the electric energy usage amount of each electricity consumption type of service, the parameter of each neutral net respectively Generated by a kind of fast algorithm；Finally, each electricity consumption type of service electric energy usage amount total electric energy usage amount of prediction it is calculated.

The technical solution used in the present invention is: a kind of intelligent grid electric energy usage amount Forecasting Methodology, including:

S1, determine electric energy usage amount characteristic parameter according to electricity consumption type of service；Described electric energy usage amount characteristic parameter represents The electric energy usage amount of corresponding electricity consumption type of service；

S2, determining the time scale of electric energy usage amount characteristic parameter, structure describes many time serieses of electric energy usage amount

Wherein,Represent the characteristic parameter original electric energy usage amount at t time point of i-th electric energy usage amount, i =1,2,3 ..., M, M represent total number of the characteristic parameter of electric energy usage amount, t express time point sequence number；

S3, according to the time scale determined in step S2, obtain the length of window during training；And it is right in window when trainingIt is normalized；

S4, according to the time scale determined in step S2, obtain history time window length；Determine the input of time point to be predicted Vector

Wherein, X_i,jT () represents when the t time point prediction i-th electric energy usage amount characteristic parameter, required input to AmountJth numerical value, T represents that history time window length, described T are the unit interval scale in a cycle time length Number；

S5, M neutral net of initialization, and determine hidden layers numbers and the hidden node number of each neutral net；By right The prediction of the i-th ' individual neutral net obtains the electricity consumption type of service electric energy usage amount that i-th electric energy usage amount characteristic parameter is corresponding； The electricity consumption type of service electric energy that the predictive value correspondence i-th electric energy usage amount characteristic parameter of described the i-th ' individual neutral net is corresponding makes The value of consumption；

S6, judge whether to need to calculate each neural network parameter group；The most then go to step S7, otherwise go to step S9；

S7, each neutral net is carried out K time training, obtain the K group parameter group that each neutral net is corresponding；

S8, judge that neutral net to be trained has completed all to train the most, the most then go to step S9；Otherwise go to step S7；

S9, neutral net is predicted, particularly as follows: by being calculated current point in time for user's future time point The predicted mean vote of the electricity consumption type of service electric energy usage amount that each electric energy usage amount characteristic parameter is corresponding；

S10, electricity consumption type of service electric energy usage amount that each electric energy usage amount characteristic parameter obtaining step S9 is corresponding Predicted mean vote carries out renormalization process, obtains the predictive value of original electric energy usage amount；

S11, the next time point of judgement, the need of being predicted, if so, go to step S2；Otherwise terminate.

Further, described step S3, with specific reference to the following formula many time serieses to electric energy usage amountReturn One change processes:

$S_i\left(t\right)=\frac{{\hat{S}}_i\left(t\right)-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}}{\max \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}}\×2-1;$

Wherein, S_iT () represents i-th electric energy usage amount characteristic parameter electricity after the normalization that the t time point is corresponding Energy usage amount,Represent the i-th electric energy usage amount characteristic parameter original electricity that in window, the jth ' individual moment is corresponding when training Can usage amount.

Further, determine that the number of hidden node in each neutral net has by sample data test described in step S5 Body is:

A1, hidden node number L of setting the i-th ' individual neutral net_i′=1；

A2, input D group training sample are trained, and calculate the error that predicts the outcome respectively, calculate the prediction of D group training sample by mistake The meansigma methods of difference；

A3, work as L_i′=L_i′When+1, calculate the meansigma methods of D group training sample forecast error；

A4, comparison current hidden node number L_i′, the meansigma methods of the D group training sample forecast error obtained and hidden node Number is L_i′When-1, whether the reduction ratio of the meansigma methods of the D group training sample forecast error obtained is less than threshold values, the most then Determine L now_i′It it is the number of hidden node in the i-th ' individual neutral net；Otherwise, step A3 is returned to.

Further, described step S6 specifically include following step by step:

The window during assessment of S61, defined parameters, during described parameter evaluation, window is by current point in time and front N₀-1 each and every one time point structure Become；

The electricity consumption type of service electric energy that during assessment of S62, defined parameters, in window, i-th electric energy usage amount characteristic parameter is corresponding makes The mean error of consumption prediction；

The time point that if each neutral net of S63 not yet initializes or has been predicted by the end of current point in time Number is less than N₀, then step S7 is performed；Otherwise go to step S64；

If the i-th class electric energy usage amount characteristic parameter correspondence electricity consumption type of service that S64 neutral net to be judged is predicted Electric energy usage amount mean error in window when parameter evaluation more than given threshold values, then goes to step S7, otherwise goes to step S9.

Further, described step S7 specifically includes following steps:

S71, input weight matrix to the i-th ' individual neutral netWith hidden node bias vectorCarry out assignment；

${\stackrel{\→}{A}}_{i^{\′}}=\left\{{\stackrel{\→}{a}}_{i^{\′},j^{\′\′}}\right\},(j^{\′\′}=1,2,3,\mathrm{...},L_{i^{\′}});$

Wherein,Represent in the i-th ' individual neutral net all input nodes to jth " the input weight of individual hidden node；

${\stackrel{\→}{b}}_{i^{\′}}=\left\{b_{i^{\′},j^{\′\′}}\right\},(j^{\′\′}=1,2,3,\mathrm{...},L_{i^{\′}});$

Wherein, b_i′,j″Represent the jth " bias that individual hidden node is corresponding in the i-th ' individual neutral net；

S72, calculating neutral net output matrix H₀；

Wherein," individual hidden node corresponds to sample to represent the i-th ' individual neutral net jth Hidden node output, t=1,2,3 ..., N, N represent training data number；It it is the input weight of the i-th ' individual neutral net；

S73, the outer power of calculating neutral net output

${\stackrel{\→}{\mathrm{\β}}}_{i^{\′}}=H_0^{+}Q;$

Wherein,Represent neutral net output matrix H₀Generalized inverse； $Q={\left[\begin{array}{c}{Y}_{i^{\′}}{\left(1\right)}^T\\ .\\ .\\ .\\ Y_{i^{\′}}{\left(N\right)}^T\end{array}\right]}_{N\×1},$ Y_i′T () is the i-th ' individual god Through network at the actual value of the t time point predicted electric energy usage amount, t=1,2,3 ..., N, N represent training data number.

Further, described step S9 specifically include following step by step:

S91, the output matrix H of the calculating each neutral net of current point in time；

$H={\left[\begin{array}{ccc}G({\stackrel{\→}{a}}_{i^{\′},1},b_{i^{\′},1},{\stackrel{\→}{X}}_{i^{\′}}(t)& \mathrm{...}& G({\stackrel{\→}{a}}_{i^{\′},L_{i^{\′}}},b_{i^{\′},L_{i^{\′}}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)\end{array}\right]}_{1\×L_{i^{\′}}};$

S92, calculate current point in time for user's future time point each electric energy usage amount characteristic parameter corresponding use electric industry Predictive value { the P of service type electric energy usage amount_i,k(t)}；

$P_{i,k}\left(t\right)=P_{i^{\′},k}\left(t\right)={\left[\begin{array}{ccc}G({\stackrel{\→}{a}}_{i^{\′},1},b_{i^{\′},1},{\stackrel{\→}{X}}_{i^{\′}}(t)& \mathrm{...}& G({\stackrel{\→}{a}}_{i^{\′},L_{i^{\′}}},b_{i^{\′},L_{i^{\′}}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)\end{array}\right]}_{1\×L_{i^{\′}}}{\left[\begin{array}{c}{\mathrm{\β}}_{i^{\′},1}\\ .\\ .\\ .\\ {\mathrm{\β}}_{i^{\′},L_{i^{\′}}}\end{array}\right]}_{L_{i^{\′}}\×1};$

Wherein, P_i′,kPredicting the outcome of the kth group parameter of (t) expression the i-th ' individual neutral net, k=1,2 ..., K；

S93, calculate current point in time for user's future time point each electric energy usage amount characteristic parameter corresponding use electric industry The predicted mean vote of service type electric energy usage amount

${\stackrel{\‾}{P}}_i\left(t\right)={\stackrel{\‾}{P}}_{i^{\′}}\left(t\right)=\frac{{\Σ}_{k=1}^{k=K}{P}_{i^{\′},k}\left(t\right)}{K};$

Wherein, ${\stackrel{\‾}{P}}_{i^{\′}}\left(t\right)$ Represent the result averaged that predicts the outcome of the K group parameter of the i-th ' individual neutral net.

Further, original electric energy usage amount described in step S10 predictive value particularly as follows:

$\begin{array}{c}{\hat{P}}_i\left(t\right)=\frac{\left(\max \right\{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\left\}\right)\left({\stackrel{\‾}{P}}_i\right(t)+1)}{2}\\ +\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\})\end{array};$

Wherein, max{*} represents and takes maximum, and min{*} represents and takes minima.

Beneficial effects of the present invention: a kind of intelligent grid electric energy usage amount Forecasting Methodology of the present invention, first according to electricity consumption The difference of user types, uses extracting data to be correlated with electric energy usage amount characteristic parameter from electric energy, line number of going forward side by side Data preprocess；So After, building multiple neutral net and predict the electric energy usage amount of each electricity consumption type of service respectively, the parameter of each neutral net is by this Shen A kind of fast algorithm that please provide generates；Finally, each electricity consumption type of service electric energy usage amount total electric energy of prediction it is calculated Usage amount；The method of the present invention takes into full account that electric energy usage amount is comprised dissimilar electricity consumption behavior and the inherence between them Space time correlation, solves the traditional prediction method restricted problem to precision of prediction, and the method for the present invention is capable of quickly, Predict exactly.

Accompanying drawing explanation

A kind of intelligent grid electric energy usage amount Forecasting Methodology block diagram that Fig. 1 provides for the present invention.

Fig. 2 for the present invention provide for predict i-th electric energy usage amount characteristic parameter corresponding electricity consumption type of service electricity The neural network structure figure of energy usage amount.

The electricity consumption type of service electric energy that prediction i-th electric energy usage amount characteristic parameter that Fig. 3 provides for the present invention is corresponding makes The neural metwork training flow chart of consumption.

Detailed description of the invention

For ease of skilled artisan understands that the technology contents of the present invention, below in conjunction with the accompanying drawings present invention is entered one Step explaination.

It is illustrated in figure 1 a kind of intelligent grid electric energy usage amount Forecasting Methodology block diagram that the present invention provides, the application relates to And the electrical network i.e. electrical network of electric energy usage amount in the power consumption of user.The technical scheme is that a kind of intelligent grid electric energy makes Consumption Forecasting Methodology, including:

S1, determine electric energy usage amount characteristic parameter according to electricity consumption type of service；Described electric energy usage amount characteristic parameter represents The electric energy usage amount of corresponding electricity consumption type of service.

Electric energy from actual intelligent grid uses monitoring device, such as intelligent electric meter, can obtain electric energy and use data.These Electric energy uses data to include: the time dependent power consumption of user, electric current, voltage, frequency etc..Different electricity consumption type of service meetings The parameters such as power consumption, electric current, voltage, frequency show its exclusive feature (such as waveform etc.), by signal analysis and place Reason, prior art can extract different electricity consumptions by analyzing the time dependent power consumption of user, electric current, voltage, frequency The time dependent power consumption of type of service.

According to the difference of electricity consumption type of service, we can use extracting data to be correlated with electric energy usage amount feature from electric energy Parameter describes the electric energy usage amount of user.In the application, the electric energy usage amount of multiple electricity consumption type of service is used to make as electric energy Consumption characteristic parameter.Concrete: describe different electricity consumption types of service, different electric energy usage amount characteristic parameters need to be used.As Residential electricity consumption user, can use time dependent electric consumption on lighting amount, kitchen socket power consumption, use in dishwasher electricity, microwave oven Power consumption, use in washing machine electricity, baking box power consumption, refrigerator power consumption, electric appliance for bathroom power consumption, muffle electric furnace power consumption, idle call Electricity and these 11 electric energy usage amount characteristic parameters of other power consumption describe its electric energy usage amount.In the application, use M Electric energy usage amount characteristic parameter describes the electric energy usage amount of user, and according to the difference of electricity consumption type of service, M can take different Value, the prediction of user's electric energy usage amount can be by realizing, the most in advance the prediction of user's difference electricity consumption type of service electric energy usage amount Survey M the characteristic parameter describing user's electric energy usage amount.

S2, determining the time scale of electric energy usage amount characteristic parameter, structure describes many time serieses of electric energy usage amount

Wherein,Represent the characteristic parameter original electric energy usage amount at t time point of i-th electric energy usage amount, i =1,2,3 ..., M, M represent total number of the characteristic parameter of electric energy usage amount, t express time point sequence number.

According to different prediction tasks, as being predicted on an hourly basis or being daily predicted, electric energy usage amount feature is joined Optional corresponding time scale U of number, such as 15 minutes, 1 hour or 1 day.After seclected time scale U, each unit interval marks In degree, the electric energy usage amount of each electricity consumption type of service is this electricity consumption type of service electric energy corresponding to electric energy usage amount characteristic parameter Usage amount, is designated asThe electricity consumption type of service corresponding to i-th electric energy usage amount characteristic parameter of i.e. t time point is former Beginning electric energy usage amount.On each time point, electric energy usage amount M the electric energy usage amount characteristic parameter of user represents, all electricity The many time serieses describing electric energy usage amount can be constituted over time, i.e. by usage amount characteristic parameterWherein, i= 1,2,3 ..., M, t are time point sequence number.

S3, according to the time scale determined in step S2, obtain the length of window during training；And it is right in window when trainingIt is normalized；

According to time scale U chosen in step S2, determine training data number N of needs, the i.e. length of window during training. In the desirable cycle time length of training data number N, such as one day, one week or January, the 50-of unit interval scale number 100 times.When being 15 minutes for U, then N is (60/15) * 24*50=4800；For U be 1 little constantly, then N is 24*100= 2400；When being 1 day for U, then N is 30*100=3000.The data of current point in time and its front N-1 time point constitute instruction Data in window when practicing.

When training in window, the available equation below many time serieses to describing electric energy usage amountCarry out normalizing Change.

$S_i\left(t\right)=\frac{{\hat{S}}_i\left(t\right)-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}}{\max \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}}\×2-1;$

Many time serieses of electric energy usage amount are normalized it can be avoided that the bigger electric energy usage amount of codomain scope is special Levy parameters distribution and fall the impact on finally predicting the outcome of the codomain scope less electric energy usage amount characteristic parameter.Hereinafter, for chatting State conveniently, S in the application_iT () represents that i-th electric energy usage amount characteristic parameter is after the normalization that the t time point is corresponding Electric energy usage amount, claimsFormer at t time point of electricity consumption type of service corresponding to i-th electric energy usage amount characteristic parameter Beginning electric energy usage amount,Represent i-th electric energy usage amount characteristic parameter jth ' individual moment corresponding former in window when the training Beginning electric energy usage amount.

S4, according to the time scale determined in step S2, obtain history time window length；Determine the input of time point to be predicted Vector

Wherein, X_{I, j}T () represents when the t time point prediction i-th electric energy usage amount characteristic parameter, required input to AmountJth numerical value, T represents that history time window length, described T are the unit interval scale in a cycle time length Number.

According to time scale U chosen in step S2, determine history time window length T.History time window length T is a cycle In time span.Such as one day, one week or January.Unit interval scale number.When being 15 minutes for U, T be (60/15) × 24=96；For U be 1 little constantly, T is 24；When being 1 day for U, T is 30.Current point in time and its front N-1 time point Data constitute history time window in data.

Owing to M electricity consumption type of service electric energy usage amount of M electric energy usage amount characteristic parameter description is at same time point There is dependency (certain of corresponding user is movable), and, each electric energy usage amount characteristic parameter is at the current value put and this electric energy Usage amount characteristic parameter value in window when history has temporal correlation.Therefore, for i-th electric energy usage amount characteristic parameter In the prediction of next time point value, i.e. for i-th electricity consumption type of service in the prediction of next time point electric energy usage amount, Value and this electric energy usage amount characteristic parameter window when history of current point in time all electric energy usage amount characteristic parameter can be used The input vector that interior all values needs as prediction.The application remembers that i-th electric energy usage amount characteristic parameter is in the next time The predictive value of point (the t+1 time point) is P_i(t) (i=1,2,3 ..., M), the input vector of its correspondence is:

${\stackrel{\→}{X}}_i\left(t\right)=\{Z_i\left(t\right),{\hat{Z}}_i\left(t\right)\},(i=1,2,3...,M).$

Wherein, Z_i(t)={ S_i(t),S_i(t-1),...,S_i(t-T+1) },P_i T () is in electric energy usage amount S of t+1 time point to electricity consumption type of service corresponding to i-th electric energy usage amount characteristic parameter_i (t+1) prediction.Y_iT () is that electricity consumption type of service that i-th electric energy usage amount characteristic parameter is corresponding is at the t+1 time point Electric energy usage amount S_i(t+1) actual value, i.e. desired value.

For hereinafter describing conveniently, the application remembers

Wherein, X_i,jT () represents when the t time point prediction i-th electric energy usage amount characteristic parameter, required input to AmountJth numerical value.Due to S_i(t) normalization, so now without again to X_i,jT () is normalized.Follow-up Using in the step that neutral net is predicted, the application will directly useInput vector as time point to be predicted.

S5, M neutral net of initialization, special by the prediction of the i-th ' individual neutral net being obtained i-th electric energy usage amount Levy the electricity consumption type of service electric energy usage amount that parameter is corresponding；And determine hidden layers numbers and the hidden node number of each neutral net.

It is illustrated in figure 2 for predicting that the electricity consumption type of service electric energy that i-th electric energy usage amount characteristic parameter is corresponding uses The neural network structure figure of amount,Row vector is tieed up, represents jth in the i-th ' individual neutral net " individual hidden node defeated for M+T-1 Enter weight；b_i′,j″It is a numerical value, represents the jth " biasing that individual hidden node is corresponding in the i-th ' individual neutral net；β_i′,j″Represent " the outer weights between individual hidden node and network output of connection jth in i-th ' individual neutral net；Represent " individual hidden node corresponds to input vector to jthI ' output, G () is activation primitive.

The electric energy usage amount of M class electricity consumption type of service is predicted by the application with using M neural network concurrent.The i-th ' The object that individual neutral net is predicted is that the electric energy of the electricity consumption type of service that i-th electric energy usage amount characteristic parameter is corresponding uses Amount, the input vector (at the t time point) of its correspondence isPredictive value is (at the t time point) P_iT (), desired value is (at the t time point) Y_i(t)。

For each neutral net, it is 1 that the application all chooses single hidden layer network, i.e. hidden layers numbers.Neural networks with single hidden layer has Have that simple in construction, training speed are fast, be not easy the advantages such as over-fitting.

Number L of hidden node in each neutral net_i′, the application is tested by sample data and determines, concrete steps are such as Under:

First hidden node number L of target setting neutral net_i′=1.

2. 50 groups of training samples of input are trained, and calculate the error that predicts the outcome respectively, calculate 50 groups of training sample predictions The meansigma methods of error.

3. make L_i′=L_i′+ 1, again calculate when using this number of hidden nodes, 50 groups of training sample forecast erroies average Value.

If L_i′During increase, the value that the forecast error meansigma methods of use training sample calculating is more last reduces ratio less than valve Value, the application selects default value to be 0.1% as threshold values, it is determined that L now_i′For in target nerve network hidden node Number；Otherwise, step 3 is returned to.

Reduce ratio particularly as follows: make current hidden node number L_i′, the meansigma methods of the D group training sample forecast error obtained For Mean1, L_i′-1 correspondence meansigma methods be Mean2, then reduce ratio be: | Mean1-Mean2 |/Mean1.

S6, judge whether to need to calculate each neural network parameter group；The most then go to step S7, otherwise go to step S9.

It is illustrated in figure 3 electricity consumption type of service electric energy usage amount corresponding to prediction i-th electric energy usage amount characteristic parameter Neural metwork training flow chart, specifically comprises the following steps that

First, during defined parameters assessment, window judges whether to need to recalculate each neural network parameter group with help.Parameter During assessment, window is by current point in time and front N thereof₀-1 each and every one time point is constituted.In this application, parameter evaluation time window length N₀Silent Recognizing value is 5.

Then, the electricity consumption type of service electric energy that during defined parameters assessment, in window, i-th electric energy usage amount characteristic parameter is corresponding The mean error of usage amount predictionFor:

${\stackrel{\‾}{\mathrm{\δ}}}_i\left(t_0\right)=\frac{{\Σ}_{t=t_0-N_0+1}^{t_0}\frac{|{\hat{P}}_i(t-1)-S_i\left(t\right)|}{S_i\left(t\right)}}{N_0}\×100\%$

Wherein, t₀For current time,By on the t-1 time point the meansigma methods of prediction electric energy usage amount, i.e. At the t-1 time point for S_iT mean predicted value that () repeatedly predicts.

Furthermore, if each neutral net not yet initializes or the time point that has been predicted by the end of current point in time Number is less than N₀, all of neutral net all enters parameter group training step, calculates the parameter group of each neutral net.Otherwise, depend on The secondary parameter group to each neutral net judges the need of recalculating.

Finally, if electricity consumption service class corresponding to the i-th electric energy usage amount characteristic parameter predicted of neutral net to be judged Type electric energy usage amount mean error in window when parameter evaluationMore than given threshold values, depending on precision of prediction requires, than Take as in the applicationFor 1.5 times of meansigma methods Mean of training sample forecast error, according to different accuracy demand, it is possible to take 2 times, 3 times, 1.2 times etc., then this neutral net needs to recalculate parameter group.Otherwise, this neutral net need not recalculate Parameter group, at the electricity consumption type of service electric energy predicting that the i-th electric energy usage amount characteristic parameter corresponding to this neutral net is corresponding During usage amount, the parameter group that can directly utilize this neutral net current is predicted.

S7, each neutral net is carried out K time training, obtain the K group parameter group that each neutral net is corresponding；The application sets Determining frequency of training is K=50, particularly as follows: be trained according to historical data, obtains the parameter group of neutral net；

Specifically include following steps:

S71, input weight matrix to the i-th ' individual neutral netWith hidden node bias vectorCarry out assignment；

Wherein,Represent that in the i-th ' individual neutral net, all input nodes are to jth The input weight of individual hidden node,

Wherein, b_i′,jRepresent that in the i-th ' individual neutral net, jth hidden node is corresponding Bias；

For predicting the neutral net of the i-th ' individual electricity consumption type of service electric energy usage amount, need its parameterWithCarry out Assignment.Wherein,L for stochastic generation_i′* the matrix of (M+T-1) size,It is the i-th ' individual In neutral net, all input nodes are to jth, and " the input weight of individual hidden node is the vector of one-dimensional a length of M+T-1.For stochastic generation L_i′* the vector of 1 size.Wherein, b_i′,j″Represent in the i-th ' individual neutral net Jth " the bias that individual hidden node is corresponding.Neutral net input weightB is biased with hidden node_i′,j″(j "=1,2, 3...,L_i′) by system stochastic generation.

Above random number can be produced at [0,1] equally distributed probability density function by one.

S72, calculating neutral net output matrix H₀；

The activation primitive of neutral net we choose Sigmoid function.

$G\left(x\right)=\frac{1}{1+e^{-x}}$

For input vectorThere is L_i′The output expression formula of the neutral net of individual hidden node is:

$Y_{i^{\′}}\left(t\right)=\underset{j^{\′\′}=1}{\overset{j^{\′\′}=L_{i^{\′}}}{\Σ}}{\mathrm{\β}}_{i^{\′},j^{\′\′}}G({\stackrel{\→}{a}}_{i^{\′},j^{\′\′}},b_{i^{\′},j^{\′\′}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)$

In formula, Y_i′(t) by the i-th ' individual neutral net at the actual value of the t time point predicted electric energy usage amount, i.e. S_i (t),It is the input weight of the i-th ' individual neutral net, b_i′,jIt is the i-th ' individual neutral net jth " biasing of individual hidden node, β_i′,jRepresent and the i-th ' individual neutral net connect jth " individual hidden node and the outer power of neutral net output," individual hidden node corresponds to sample to represent the i-th ' individual neutral net jthHidden node defeated Go out.For the hidden node of addition type,Expression formula be:

$G({\stackrel{\→}{a}}_{i^{\′},j^{\′\′}},b_{i^{\′},j^{\′\′}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)=\frac{1}{1+\exp (-({\stackrel{\→}{a}}_{i^{\′},j^{\′\′}}\·{\stackrel{\→}{X}}_{i^{\′}}\left(t\right)+b_{i^{\′},j^{\′\′}}\left)\right)}$

Wherein,Represent weight vectorsAnd input vectorInner product.G () defined herein: R- > R is one and is named as G, definition territory, codomain are the mapping of real number；The definition of function G () is i.e. function aboveI.e." " represents this function All independent variables；In " R-> R ", previous R represents that definition territory is real number, and later R represents that codomain is real number,-> for mapping symbol Number.

Use the input data of N (N is the training data number defined in step 3) individual continuous time point The output matrix H of the i-th ' individual neutral net can be calculated₀:

S73, the outer power of calculating neutral net output

Still using the N, N described in step S72 is the training data number defined in step S3, individual continuous time point Input dataCalculate the i-th ' outer power of individual neutral net output's Computing formula is as follows:

${\stackrel{\→}{\mathrm{\β}}}_{i^{\′}}=H_0^{+}Q;$

Wherein,Represent neutral net output matrix H₀Generalized inverse, Defeated for neutral net Go out matrix H₀Transposition,For matrixInverse matrix；Y_i′T () is the i-th ' individual neutral net At the actual value of the t time point predicted electric energy usage amount, t=1,2,3 ..., N, N represent training data number.It it is one L_i′The column vector of × 1.

S8, judge that neutral net to be trained has completed all to train the most, the most then go to step S9；Otherwise go to step S7.Successively each neutral net to be trained is judged, if neutral net to be judged is not fully complete 50 stand-alone trainings, then turn Training is proceeded to step S7；When being needed to train neutral net the most to complete 50 stand-alone trainings, then complete and carry out The neural network parameter group that the prediction of current point in time electric energy usage amount needs calculates.

One group of parameter of the i-th ' individual neutral net can be obtained by step S7WithTraining nerve is treated to each Network, the application uses identical input dataCarry out 50 times Stand-alone training.

Successively each neutral net to be trained is judged.If neutral net to be judged is not fully complete 50 independent instructions Practice, then forward step S7 to and proceed training.When being needed to train neutral net the most to complete 50 stand-alone trainings, then complete The neural network parameter group that carrying out the prediction of current point in time electric energy usage amount needs calculates, then go to step S9.

S9, neutral net is predicted, particularly as follows: by being calculated current point in time for user's future time point The predicted mean vote of the electricity consumption type of service electric energy usage amount that each electric energy usage amount characteristic parameter is corresponding；Specifically include following substep Rapid:

S91, the output matrix H of the calculating each neutral net of current point in time；

Often group parameter for the i-th ' individual neutral netWithIts output matrix can calculate by equation below:

$H={\left[\begin{array}{ccc}G({\stackrel{\→}{a}}_{i^{\′},1},b_{i^{\′},1},{\stackrel{\→}{X}}_{i^{\′}}(t)& \mathrm{...}& G({\stackrel{\→}{a}}_{i^{\′},L_{i^{\′}}},b_{i^{\′},L_{i^{\′}}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)\end{array}\right]}_{1\×L_{i^{\′}}};$

S92, calculate current point in time for user's future time point each electric energy usage amount characteristic parameter corresponding use electric industry Predictive value { the P of service type electric energy usage amount_i,k(t)}；

Kth group parameter for the i-th ' individual neutral netWithAll can get an i-th class electric energy usage amount special Levy the predictive value P of electricity consumption type of service electric energy usage amount corresponding to parameter_i,kT (), by calculating the kth of the i-th ' individual neutral net Predicting the outcome of group parameter, correspondence obtains P_i,k(t), computing formula is as follows:

$P_{i,k}\left(t\right)=P_{i^{\′},k}\left(t\right)={\left[\begin{array}{ccc}G({\stackrel{\→}{a}}_{i^{\′},1},b_{i^{\′},1},{\stackrel{\→}{X}}_{i^{\′}}(t)& \mathrm{...}& G({\stackrel{\→}{a}}_{i^{\′},L_{i^{\′}}},b_{i^{\′},L_{i^{\′}}},{\stackrel{\→}{X}}_{i^{\′}}(t\left)\right)\end{array}\right]}_{1\×L_{i^{\′}}}{\left[\begin{array}{c}{\mathrm{\β}}_{i^{\′},1}\\ .\\ .\\ .\\ {\mathrm{\β}}_{i^{\′},L_{i^{\′}}}\end{array}\right]}_{L_{i^{\′}}\×1}$

S93, calculate current point in time for user's future time point each electric energy usage amount characteristic parameter corresponding use electric industry The predicted mean vote of service type electric energy usage amountTie by calculating the prediction of the K=50 group parameter of the i-th ' individual neutral net The result that fruit is averaged, correspondence obtainsThus obtain

${\stackrel{\‾}{P}}_i\left(t\right)={\stackrel{\‾}{P}}_{i^{\′}}\left(t\right)=\frac{{\Σ}_{k=1}^{k=50}{P}_{i^{\′},k}\left(t\right)}{50}.$

S10, electricity consumption type of service electric energy usage amount that each electric energy usage amount characteristic parameter obtaining step S9 is corresponding Predicted mean vote carries out renormalization process, obtains the predictive value of original electric energy usage amount；

The predicted mean vote of each electricity consumption type of service electric energy usage amount is carried out renormalization, obtains each electricity consumption type of service The predictive value of original electric energy usage amount

The current point in time predictive value to user's future time point each electricity consumption type of service original electric energy usage amountCan Use equation below calculates:

$\begin{array}{c}{\hat{P}}_i\left(t\right)=\frac{\left(\max \right\{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\}-\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\left\}\right)\left({\stackrel{\‾}{P}}_i\right(t)+1)}{2}\\ +\min \{{\hat{S}}_i\left(j^{\′}\right),j^{\′}=1,2,\mathrm{3...},N\})\end{array}$

S11, the next time point of judgement, the need of being predicted, if so, go to step S2；Otherwise terminate.

Those of ordinary skill in the art it will be appreciated that embodiment described here be to aid in reader understanding this Bright principle, it should be understood that protection scope of the present invention is not limited to such special statement and embodiment.For ability For the technical staff in territory, the present invention can have various modifications and variations.All within the spirit and principles in the present invention, made Any modification, equivalent substitution and improvement etc., within should be included in scope of the presently claimed invention.

Claims (7)

1. an intelligent grid electric energy usage amount Forecasting Methodology, it is characterised in that including:

S1, determine electric energy usage amount characteristic parameter according to electricity consumption type of service；Described electric energy usage amount characteristic parameter represents corresponding The electric energy usage amount of electricity consumption type of service；

S2, determining the time scale of electric energy usage amount characteristic parameter, structure describes many time serieses of electric energy usage amount

Wherein,Represent the characteristic parameter of i-th electric energy usage amount in the original electric energy usage amount of t time point, i=1, 2,3 ..., M, M represent total number of the characteristic parameter of electric energy usage amount, t express time point sequence number；

S3, according to the time scale determined in step S2, obtain the length of window during training；And it is right in window when trainingEnter Row normalized；

S4, according to the time scale determined in step S2, obtain history time window length；Determine the input vector of time point to be predicted

Wherein, X_i,jT () represents when the t time point predicts i-th electric energy usage amount characteristic parameter, required input vectorJth numerical value, T represents that history time window length, described T are the unit interval scale in a cycle time length Number；

S5, M neutral net of initialization, and determine hidden layers numbers and the hidden node number of each neutral net；By to the i-th ' The prediction of individual neutral net obtains the electricity consumption type of service electric energy usage amount that i-th electric energy usage amount characteristic parameter is corresponding；Described The electricity consumption type of service electric energy usage amount that the predictive value correspondence i-th electric energy usage amount characteristic parameter of the i-th ' individual neutral net is corresponding Value；

S6, judge whether to need to calculate each neural network parameter group；The most then go to step S7, otherwise go to step S9；

S7, each neutral net is carried out K time training, obtain the K group parameter group that each neutral net is corresponding；

S8, judge that neutral net to be trained has completed all to train the most, the most then go to step S9；Otherwise go to step S7；

S9, neutral net is predicted, particularly as follows: by being calculated current point in time for user's future time each electricity of point The predicted mean vote of the electricity consumption type of service electric energy usage amount that energy usage amount characteristic parameter is corresponding；

S10, the prediction of the electricity consumption type of service electric energy usage amount that each electric energy usage amount characteristic parameter obtaining step S9 is corresponding Meansigma methods carries out renormalization process, obtains the predictive value of original electric energy usage amount；

S11, the next time point of judgement, the need of being predicted, if so, go to step S2；Otherwise terminate.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that described step S3, with specific reference to the following formula many time serieses to electric energy usage amountIt is normalized:

Wherein, S_iT () represents the electric energy use after the normalization that the t time point is corresponding of the i-th electric energy usage amount characteristic parameter Amount,Represent that the i-th electric energy usage amount characteristic parameter original electric energy that in window, the jth ' individual moment is corresponding when training uses Amount.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that institute in step S5 State by sample data test determine hidden node in each neutral net number particularly as follows:

A1, hidden node number L of setting the i-th ' individual neutral net_i′=1；

A2, input D group training sample are trained, and calculate the error that predicts the outcome respectively, calculate D group training sample forecast error Meansigma methods；

A3, work as L_i′=L_i′When+1, calculate the meansigma methods of D group training sample forecast error；

A4, comparison current hidden node number L_i′, the meansigma methods of the D group training sample forecast error obtained and hidden node number For L_i′When-1, whether the reduction ratio of the meansigma methods of the D group training sample forecast error obtained is less than threshold values, if, it is determined that L now_i′It it is the number of hidden node in the i-th ' individual neutral net；Otherwise, step A3 is returned to.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that described step S6 Specifically include following step by step:

The window during assessment of S61, defined parameters, during described parameter evaluation, window is by current point in time and front N₀-1 each and every one time point is constituted；

The electricity consumption type of service electric energy usage amount that during assessment of S62, defined parameters, in window, i-th electric energy usage amount characteristic parameter is corresponding The mean error of prediction；

If the time point number that each neutral net of S63 not yet initializes or has been predicted by the end of current point in time is little In N₀, then step S7 is performed；Otherwise go to step S64；

If the electric energy of the i-th class electric energy usage amount characteristic parameter correspondence electricity consumption type of service that S64 neutral net to be judged is predicted Usage amount mean error in window when parameter evaluation more than given threshold values, then goes to step S7, otherwise goes to step S9.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that described step S7 Specifically include following steps:

S71, input weight matrix to the i-th ' individual neutral netWith hidden node bias vectorCarry out assignment；

Wherein,Represent in the i-th ' individual neutral net all input nodes to jth " the input weight of individual hidden node；

Wherein, b_i′,j″Represent the jth " bias that individual hidden node is corresponding in the i-th ' individual neutral net；

S72, calculating neutral net output matrix H₀；

Wherein," individual hidden node corresponds to sample to represent the i-th ' individual neutral net jthHidden layer Node exports, t=1, and 2,3 ..., N, N represent training data number；It it is the input weight of the i-th ' individual neutral net；

S73, the outer power of calculating neutral net output

Wherein,Represent neutral net output matrix H₀Generalized inverse；Y_i′T () is the i-th ' individual neutral net At the actual value of the t time point predicted electric energy usage amount, t=1,2,3 ..., N, N represent training data number.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that described step S9 Specifically include following step by step:

S91, the output matrix H of the calculating each neutral net of current point in time；

S92, calculate current point in time for electricity consumption service class corresponding to user's future time point each electric energy usage amount characteristic parameter Predictive value { the P of type electric energy usage amount_i,k(t)}；

Wherein, P_i′,kPredicting the outcome of the kth group parameter of (t) expression the i-th ' individual neutral net, k=1,2 ..., K；

S93, calculate current point in time for electricity consumption service class corresponding to user's future time point each electric energy usage amount characteristic parameter The predicted mean vote of type electric energy usage amount

Wherein,Represent the result averaged that predicts the outcome of the K group parameter of the i-th ' individual neutral net.

A kind of intelligent grid electric energy usage amount Forecasting Methodology the most according to claim 1, it is characterised in that step S10 institute State the predictive value of original electric energy usage amount particularly as follows:

Wherein, max{*} represents and takes maximum, and min{*} represents and takes minima.