CN106651020A - Short-term power load prediction method based on big data reduction - Google Patents

Abstract

The invention provides a short-term power load prediction method based on big data reduction. The method comprises the steps that first, the Lasso principle is utilized to remove redundant data and bad data in big data, and then environment factor variables are subjected to dimensionality reduction and feature extraction through principal component analysis (PCA); extracted feature vectors and historical load data obtained after simple processing are jointly used as input of an Elman neural network to perform training prediction. Through the method, prediction precision and speed of a short-term power load are obviously improved.

Description

A kind of Short-Term Load Forecasting Method brief based on big data

Technical field

The present invention relates to Techniques for Prediction of Electric Loads field, more particularly to a kind of short term power brief based on big data Load forecasting method.

Background technology

Load forecast is one of important process of power supply department, is to ensure power system reliable power supply and safe operation Premise.Accurately load prediction can economically arrange the start and stop of electrical network internal generator group, accomplish to improve economic effect Benefit and social benefit.In the face of the fast development of nowadays intelligent grid, electric load influence factor increases, and data exponentially increase It is long, progressively constitute big data it is multidimensional the characteristics of, traditional data analysis pattern cannot meet demand.How to accomplish efficiently Accurately predicting the electric load of this feature becomes the key issue of present solution.Current Short-term Load Forecasting In, it is that dynamic network is recognized using static feedforward network using wide BP neural network, dynamic time is modeled Problem is changed into Static-state Space modeling problem so that precision of prediction is relatively low, and needs substantial amounts of sample data when training, and makes Obtain predetermined speed also relatively slow, this is all significantly increased the operating cost for causing electric power, i.e., in the face of the big data feature of electric load, Current short-term load forecasting method can not fully meet the demand of intelligent grid.Therefore, we have proposed a kind of based on big The Short-Term Load Forecasting Method of data reduction.

The content of the invention

Present invention aim at providing the precision of prediction and speed that short-term electric load is significantly improved under a kind of intelligent grid The Short-Term Load Forecasting Method brief based on big data.

For achieving the above object, technical scheme below is employed, the inventive method is comprised the following steps：

Step 1, selects the sampling point load sequence on n same type date before prediction day, the sampled point on each date For 48 points, i.e., sample once per 30min；

Step 2, obtains the impact load associated weather factor data for selecting n same type date and prediction day；

Step 3, (Lasso) principle is shunk by the raw power load for obtaining and associated weather factor using least absolute value It is brief that big data carries out high dimensional data, rejects bad data, obtains useful load sequence；

Weather environment variable factors are carried out dimensionality reduction and feature extraction by step 4 by principal component analysis (PCA), obtain ring Border characterization factor；

Step 5, sets up Elman dynamic neural network Short-term Load Forecastings, by extract characteristic vector and Historical load data is trained prediction collectively as the input of Elman neutral nets, and by Elman methods (n+1)th day is predicted Each moment load value.

Associated weather factor loading data in the step 2 include：Fine day (X1), cloudy (X2), rainy day (X3), highest Temperature (X4), the lowest temperature (X5), air pressure (X6), humidity (X7), radiation (X8), wind speed (X9), cloud amount (X10) etc. 10, this The weather conditions data of 10 correlations will be used as emulation data input；

The step 3) in least absolute value shrink (Lasso) principle it is as follows：

Data mining is carried out to load big data using Lasso algorithms, reject redundant data, so as to for Load Forecast Algorithm Brief and effective characteristic is provided；Lasso methods are a kind of Shrinkage estimations, by one penalty function of construction obtain one compared with For the model of refining so that it compresses some coefficients, and it is zero to concurrently set some coefficients, therefore remain the spy of subset contraction Point；

It is provided with linear regression model (LRM)：

Y=alpha+betas₁x₁+β₂x₂+…+β_px_p+ε (1)

In formula, α is constant term；β₁,β₂,…β_pFor regression coefficient；ε is Disturbance；(x_i1,x_i2,...,x_ip；y_i), i =1,2 ..., n is the n group observationses of variable, needs to meetWherein j=1,2 ..., p；

The Lasso of constant term and regression coefficient estimates to be defined as：

The detailed process of Data Dimensionality Reduction is as follows：

(a) constraints：S is penalty；

B () makesRepresent β_jLeast-squares estimation, then have

C () constantly increases when s values, the data into regression model will increase, and when certain value is reached, all data are all Regression model can be entered；When s values reduce to a certain extent, the estimate of some regression coefficients is 0, and model proposes that coefficient is 0 Variable, so as to reach the purpose of dimensionality reduction.

In step 4, principal component analysis main purpose is, to meteorological data dimension-reduction treatment in load forecast, to extract many days Gas factor characteristic index, with historical load data collectively as modeling object, the characteristic quantity for making foundation had both characterized each factor comprehensively Impact to electric load, and can Simplified prediction model；

N sample is suppose there is, each sample has p variable, constitute the data matrix of n × p rank

Detailed process is as follows：

(a) data normalization --- initial data criterion is turned to into the valid data between [0,1]；

B () calculates correlation matrix

In formula, (i, j=1,2 ..., are p) coefficient correlation between former variable xi and xj to rij, and its computing formula is：

Because R is real symmetric matrix (i.e. rij=rji), triangle element or lower triangle element need to be only calculated；

C () calculates eigen vector

| the λ I-R |=0 that solve characteristic equation first, generally obtain eigenvalue λ with Jacobi method_i(i=1,2 ..., p), and make Its order arrangement by size, i.e. λ₁≥λ₂≥…,≥λ_p≥0；Then obtain respectively corresponding to eigenvalue λ_iCharacteristic vector a_i(i =1,2 ..., p)；

D () calculates principal component contributor rate and contribution rate of accumulative total

Overall merit and weighted sum are carried out to m principal component

F=w₁z₁+w₂z₂+…+w_mz_m (7)

In formula, wi is the contribution rate of principal component, and computing formula is：

Representated by contribution rate is the percentage of original index information content shared by i-th principal component, so first principal component Should account for the largest percentage, then gradually successively decrease.Front m principal component adds up variance contribution ratio：

In order to reach the purpose of dimensionality reduction, the contribution rate of accumulative total of current m principal component reaches more than 85%, it is possible to use front m Individual principal component is substituting p original evaluation index；

The construction of (e) new samples matrix:Definition：Note x1, x2 ..., xP are former variable index, z1, z2 ..., zm (m≤p) For new variables index, each sample value of each principal component is calculated according to formula (12) and formula (13)；

In step 5, concrete Forecasting Methodology is as follows：

Elman neutral nets are to increase a undertaking layer in the feedforward network hidden layer of BP, as a step time delay operator, To reach the purpose of memory, by storing internal state so as to possess the function of dynamic characteristic, so that system has adapting to The characteristic of catastrophic event；

The non-linear state space expression of Elman networks is:

In formula, k is the number of times of neural metwork training；Y is that n ties up output vector；X is hidden neuron output vector；U is defeated Incoming vector；Xc is feedback state vector；W3 is intermediate layer to output layer connection weight；W2 is input layer to intermediate layer connection weight Value；W1 is the connection weight for accepting layer to intermediate layer.G is the transmission function of output neuron, is linear group of intermediate layer output Close；F is the transmission function of hidden neuron, frequently with s functions；

Elman networks are also adopted by BP algorithm and carry out modified weight, and study target function adopts sum of squared errors function：

In formula,For target input vector.

Compared with prior art, the inventive method has the advantage that：

1st, redundant data and bad data in intelligent grid big data can be rejected using Lasso principles so as to number According to more effectively brief；

2nd, being introduced into Lasso-PCA carries out dimension-reduction treatment to the meteorological data in load forecast, extracts many weather conditions Characteristic quantity, with historical load data collectively as modeling object, the characteristic quantity for making foundation had both characterized each factor to power load comprehensively The impact of lotus, and prediction input data can be simplified, while using dynamic Elman neural network prediction models, hence it is evident that improve electric power The accuracy and speed of load prediction.

Description of the drawings

Fig. 1 is the schematic flow sheet of the inventive method.

Fig. 2 is the Elman neural network structure figures of the inventive method.

Fig. 3 is the result figure that the inventive method carries out principal component analysis to 10 environmental variance factors.

Fig. 4 is the comparison diagram of the prediction load curve under distinct methods and realized load curve.

Fig. 5 is the predicated error comparison diagram under distinct methods.

Specific embodiment

Below in conjunction with the accompanying drawings the present invention will be further described：

As shown in figure 1, the present invention provides a kind of Short-Term Load Forecasting Method brief based on big data, including it is following Step：

1) the load sequence of (sampling once per 30min) on n same type date a few days ago is predicted in selection at 48 points；

2) the impact load factor data for selecting n same type date and prediction day are obtained；

3) using least absolute value (Lasso) principle is shunk by the raw power load for obtaining and the big number of associated weather factor According to carrying out, high dimensional data is brief, rejects bad data, obtains useful data collection；

4) dimensionality reduction and feature extraction are carried out to environmental factor variable by principal component analysis (PCA), obtain environmental characteristic because Son；

5) Elman dynamic neural network Short-term Load Forecastings are set up, by the characteristic vector extracted and history Load data is trained prediction collectively as the input of Elman neutral nets.

As shown in Fig. 2 Elman neutral nets are generally divided into 4 layers：Input layer, intermediate layer (hidden layer) accepts layer and output Layer.The connection of input layer, hidden layer and output layer is similar to feedforward network, and the unit of input layer only plays signal transmission effect, defeated Go out the linear weighting effect of layer unit.The transmission function of implicit layer unit can be using linearly or nonlinearly function, and accepting layer is used for Memory hidden layer past state, and subsequent time in conjunction with network inputs together as implicit layer unit input so that network have There is dynamic memory function, so as to reach the purpose of dynamic modeling.

Application example：

It is this regional daily load to be carried out according to setting up PCA-Elman forecast models with certain Power system load data for saving Prediction, selects Power system load data (y) and fine day (X1), cloudy (X2), the rain in this regional 1 day～March 8 March in 2004 My god (X3), the highest temperature (X4), the lowest temperature (X5), air pressure (X6), humidity (X7), radiation (X8), wind speed (X9), cloud amount Etc. (X10) the environmental factor data of 10 correlations are used as emulation data.During prediction, using the data of first 7 days as training sample, often Used as input vector, the load sequence of the 4th day is obtained 4 groups of training to the load sequence of first 3 days as object vector, according to this rule Sample carries out sample training to this model, finally the data of the 8th day is verified into the standard of network model as the test sample of network True property.

Prediction process is carried out according to flow chart described in Fig. 1.Bad data is carried out to being input into load sequence using Lasso principles Process, reach input data accurately brief, then the meteorological data in load forecast is carried out at dimensionality reduction using PCA Reason, extracts many weather conditions characteristic quantities, with it is brief after historical load data collectively as modeling object, be input to Elman refreshing So as to carrying out load prediction in Jing network models.Elman neural network structure figures are as shown in Figure 2.

As Fig. 3 be the result figure that 10 environmental variance factors are carried out with principal component analysis, by principal component characteristic value by greatly to Minispread, obtains the accumulative variance contribution ratio of front 3 principal components and can reach 85% and this concludes the description of front 3 principal components almost to include The information that whole influence factors have, the new factor, choosing are chosen according to contribution rate more than the principle of 85% (characteristic value is more than 1) Select front 3 characteristic values and calculate corresponding characteristic vector, as the input of dynamic neural network.

The comparison diagram of the prediction load curve and the realized load curve that are respectively as shown in Figure 4, Figure 5 under distinct methods and respectively Predicated error comparison diagram under method, it is more accurate when being predicted to short-term electric load compared with other Forecasting Methodologies by the method True property.

As described above, the method is rejected redundant data in big data and bad data first with Lasso principles, so Afterwards dimensionality reduction and feature extraction are carried out to environmental factor variable by principal component analysis PCA.The characteristic vector of extraction and brief place Historical load data after reason is trained prediction collectively as the input of Elman neutral nets, finally gives and predicts before Method is compared and accurately predicted the outcome.The proposition of the method is provided for the power-system short-term load forecasting under intelligent grid A kind of new approaches.

Embodiment described above is only that the preferred embodiment of the present invention is described, not to the model of the present invention Enclose and be defined, on the premise of without departing from design spirit of the present invention, technical side of the those of ordinary skill in the art to the present invention Various modifications and improvement that case is made, all should fall in the protection domain of claims of the present invention determination.

Claims (5)

1. a kind of Short-Term Load Forecasting Method brief based on big data, it is characterised in that methods described includes following step Suddenly：

Step 1, selects the sampling point load sequence on n same type date before prediction day；The sampled point on each date is 48 Point, i.e., sample once per 30min；

Step 2, obtains the impact load associated weather factor data for selecting n same type date and prediction day；

Step 3, (Lasso) principle is shunk by the raw power load for obtaining and the big number of associated weather factor using least absolute value According to carrying out, high dimensional data is brief, rejects bad data, obtains useful load sequence；

Weather environment variable factors are carried out dimensionality reduction and feature extraction by step 4 by principal component analysis (PCA), obtain environment special Levy the factor；

Step 5, sets up Elman dynamic neural network Short-term Load Forecastings, by the characteristic vector extracted and history Load data is trained prediction collectively as the input of Elman neutral nets, and by Elman methods (n+1)th day each is predicted The load value at moment.

2. a kind of Short-Term Load Forecasting Method brief based on big data according to claim 1, it is characterised in that： The associated weather factor loading data include fine day X1, cloudy day X2, rainy day X3, highest temperature X4, lowest temperature X5, air pressure X6, humidity X7, radiation X8, wind speed X9, cloud amount X10, the weather conditions data of above-mentioned 10 correlations will be used as emulation data input.

3. a kind of Short-Term Load Forecasting Method brief based on big data according to claim 1, it is characterised in that In step 3, the concrete grammar that least absolute value shrinks (Lasso) principle is as follows：

Data mining is carried out to load big data using Lasso algorithms, redundant data is rejected, so as to provide for Load Forecast Algorithm Brief and effective characteristic；Lasso methods are a kind of Shrinkage estimations, and by one penalty function of construction a more essence is obtained The model of refining so that it compresses some coefficients, it is zero to concurrently set some coefficients, therefore the characteristics of remain subset and shrink；

It is provided with linear regression model (LRM)：

Y=alpha+betas₁x₁+β₂x₂+…+β_px_p+ε (1)

In formula, α is constant term；β₁,β₂,…β_pFor regression coefficient；ε is Disturbance；(x_i1,x_i2,...,x_ip；y_i), i=1, 2 ..., n is the n group observationses of variable, needs to meetWherein j=1,2 ..., p；

The Lasso of constant term and regression coefficient estimates to be defined as：

$(\widehat{\mathrm{\α}},\widehat{\mathrm{\β}})=\arg \min \left\{\underset{i=1}{\overset{n}{\Σ}}{(y_i-\mathrm{\α}-\underset{i=1}{\overset{p}{\Σ}}{\mathrm{\β}}_j{x}_{ij})}^2\right\}---\left(2\right)$

The detailed process of Data Dimensionality Reduction is as follows：

(a) constraints：S is penalty；

B () makesRepresent β_jLeast-squares estimation, then have

C () constantly increases when s values, the data into regression model will increase, and when certain value is reached, all data all can be entered Enter regression model；When s values reduce to a certain extent, the estimate of some regression coefficients is 0, and model proposes the variable that coefficient is 0, So as to reach the purpose of dimensionality reduction.

4. a kind of Short-Term Load Forecasting Method brief based on big data according to claim 1, it is characterised in that Step 4 principal component analysis main purpose is, to meteorological data dimension-reduction treatment in load forecast, to extract many weather conditions features Index, with historical load data collectively as modeling object, the characteristic quantity for making foundation had both characterized each factor to electric load comprehensively Impact, and can Simplified prediction model；

N sample is suppose there is, each sample has p variable, constitute the data matrix of n × p rank

$X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{...}& x_{1p}\\ x_{21}& x_{22}& \mathrm{...}& x_{2p}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ x_{n1}& x_{n2}& \mathrm{...}& x_{np}\end{array}\right]---\left(3\right)$

Detailed process is as follows：

(a) data normalization --- initial data criterion is turned to into the valid data between [0,1]；

$\hat{x}=\frac{x-x_{\min }}{x_{max}-x_{\min }}---\left(4\right)$

B () calculates correlation matrix

$R=\left[\begin{array}{cccc}{r}_{11}& r_{12}& \mathrm{...}& r_{1p}\\ r_{21}& r_{22}& \mathrm{...}& r_{2p}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ r_{p1}& r_{p2}& \mathrm{...}& r_{pp}\end{array}\right]---\left(5\right)$

In formula, (i, j=1,2 ..., are p) coefficient correlation between former variable xi and xj to rij, and its computing formula is：

$r_{ij}=\frac{\underset{k=1}{\overset{n}{\Σ}}(x_{ki}-{\stackrel{\‾}{x}}_i)(x_{kj}-{\stackrel{\‾}{x}}_j)}{\sqrt{\underset{k=1}{\overset{n}{\Σ}}{(x_{ki}-{\stackrel{\‾}{x}}_i)}^2\underset{k=1}{\overset{n}{\Σ}}{(x_{kj}-{\stackrel{\‾}{x}}_j)}^2}}---\left(6\right)$

Because R is real symmetric matrix (i.e. rij=rji), triangle element or lower triangle element need to be only calculated；

C () calculates eigen vector

| the λ I-R |=0 that solve characteristic equation first, generally obtain eigenvalue λ with Jacobi method_i(i=1,2 ..., p), and make it by big Little order arrangement, i.e. λ₁≥λ₂≥…,≥λ_p≥0；Then obtain respectively corresponding to eigenvalue λ_iCharacteristic vector a_i(i=1, 2,…,p)；

D () calculates principal component contributor rate and contribution rate of accumulative total

Overall merit and weighted sum are carried out to m principal component

F=w₁z₁+w₂z₂+…+w_mz_m (7)

In formula, wi is the contribution rate of principal component, and computing formula is：

$w_i=\frac{{\mathrm{\λ}}_i}{\underset{j=1}{\overset{n}{\Σ}}{\mathrm{\λ}}_j}---\left(8\right)$

Representated by contribution rate is the percentage of original index information content shared by i-th principal component, so first principal component should Account for the largest percentage, then gradually successively decrease；Front m principal component adds up variance contribution ratio：

$\mathrm{\ρ}=\underset{i}{\overset{m}{\Σ}}{\mathrm{\λ}}_i/\underset{j}{\overset{n}{\Σ}}{\mathrm{\λ}}_j---\left(9\right)$

In order to reach the purpose of dimensionality reduction, the contribution rate of accumulative total of current m principal component reaches more than 85%, it is possible to front m master Composition is substituting p original evaluation index；

The construction of (e) new samples matrix:Definition：Note x1, x2 ..., xP are former variable index, z1, z2 ..., and zm (m≤p) is new Variable index, according to formula (12) and formula (13) each sample value of each principal component is calculated；

$\left\{\begin{array}{c}{z}_1=a_{11}{x}_1+a_{12}{x}_2+...+a_{1p}{x}_p\\ z_2=a_{21}{x}_1+a_{22}{x}_2+...+a_{2p}{x}_p\\ \mathrm{..........}\\ z_m=a_{m1}{x}_1+a_{m2}{x}_2+...+a_{mp}{x}_p\end{array}\right.---\left(10\right)$

$a_{i1}^2+...+a_{ip}^2=1---\left(11\right).$

5. a kind of Short-Term Load Forecasting Method brief based on big data according to claim 1, it is characterised in that In step 5, concrete Forecasting Methodology is as follows：

Elman neutral nets are to increase a undertaking layer in the feedforward network hidden layer of BP, as a step time delay operator, to reach To the purpose of memory, by storing internal state so as to possess the function of dynamic characteristic, so that system has aristogenesis The characteristic of event；

The non-linear state space expression of Elman networks is:

$\left\{\begin{array}{c}y\left(k\right)=g\left(w^{3}x\right(k\left)\right)\\ x\left(k\right)=f\left(w^1{x}_c\right(k)+w^{2}u(k-1\left)\right)\\ x\left(c\right)=x(k-1)\end{array}\right.---\left(12\right)$

In formula, k is the number of times of neural metwork training；Y is that n ties up output vector；X is hidden neuron output vector；U for input to Amount；Xc is feedback state vector；W3 is intermediate layer to output layer connection weight；W2 is input layer to intermediate layer connection weight；w1 To accept layer to the connection weight in intermediate layer；G is the transmission function of output neuron, is the linear combination of intermediate layer output；F is The transmission function of hidden neuron, frequently with s functions；

Elman networks are also adopted by BP algorithm and carry out modified weight, and study target function adopts sum of squared errors function：

$E\left(w\right)=\underset{k=1}{\overset{n}{\Σ}}{\left(y_k\right(k)-{\stackrel{\‾}{y}}_k(w\left)\right)}^2---\left(13\right)$

In formula,For target input vector.