CN106022954A - Multiple BP neural network load prediction method based on grey correlation degree - Google Patents

Abstract

The invention discloses a multiple BP neural network load prediction method based on grey correlation degree. The method comprises the following steps: 1) carrying out load sequence correlation analysis based on grey correlation degree; 2) carrying out clustering based on a shortest distance method to determine a member set of a multiple BP neural network; 3) determining multiple number of the multiple BP neural network based on effectiveness index; 4) improving the BP neural network by introducing a momentum factor and adopting a multiple-calculation averaging method to solve the problem of easy local convergence of the BP neural network, and improve anti-vibration capability thereof; and 5) carrying out short-term power load prediction through an established multiple BP neural network prediction model. The method solves the problem of easy local convergence of the BP neural network and improves anti-vibration capability thereof; and compared with a conventional BP neural network prediction model, the multiple BP neural network has a better prediction effect.

Description

Multi-BP Neural Networks load forecasting method based on grey relational grade

Technical field

The present invention relates to the technical field of power-system short-term load forecasting application, particularly to a kind of many based on grey relational grade Weight BP neutral net load forecasting method.

Background technology

Load forecast ensure Power System Planning with reliably, have in terms of economical operation and be of great significance.Along with existing Going deep into of the most progressive and intelligent grid of generation technique, load prediction theory and technology grows a lot.For many years, power load Lotus Forecasting Methodology and theory continue to bring out, time series method, fuzzy theory, regression analysis, regression support vector machine, pattra leaves The technology such as this and neutral net are that load forecast provides good technical support.But existing algorithm still suffers from certain limitation Property.Time series method: higher to historical data accuracy, during short-term load forecasting, insensitive to weather conditions, it is difficult to solve The problem that the short-term load forecasting precision that causes because of meteorological factor is the highest.Regression analysis: from statistical average meaning visual angle quantitatively Describe the quantitative relation between observed variable, but limited bigger by load data gauge mould.Regression support vector machine: the party Method has good generalization ability, but can be because of the optimizing of the γ-value parameter to penalty coefficient c, the e of loss function and kernel function Cause the training time the most tediously long, especially when training sample set is larger, embody the most prominent.

In view of BP neutral net, there is the strongest non-linear mapping capability, self-learning capability and fault-tolerant ability, be applied to During load small data set, there is the advantages such as precision of prediction is higher, training speed is fast.But traditional BP Application of Neural Network is in negative During lotus prediction, still suffering from a key issue, along with the increase of load sample quantity, neural network forecast precision may decline, i.e. So-called " over-fitting " problem, and when can cause neural metwork training, convergence rate is slack-off simultaneously.Its reason is historical load All there is peak-valley difference in the load of data every day, and fluctuation is relatively big, and therefore whole load datas directly share a BP nerve Network.Obviously, when training, network is the training error emphasizing entirety, it will the problem " over-fitting " occur, can cause During the prediction of later stage actual load, generalization ability is more weak, and along with the increase of training sample, load prediction speed also will be decreased obviously.

Summary of the invention

The technical problem to be solved is to provide a kind of multi-BP Neural Networks load prediction side based on grey relational grade Method, causes the more weak problem of generalization ability for traditional BP Application of Neural Network in load prediction because there is " over-fitting ", Based on grey relational grade and knearest neighbour method, definition characterizes the Validity Index that cluster is good and bad, determines the reasonable of forecast model with this Tuple.

For solving above-mentioned technical problem, the technical solution used in the present invention is:

A kind of multi-BP Neural Networks load forecasting method based on grey relational grade, comprises the following steps: step one: use The relatedness of load sequence is analyzed by grey relevant degree method；Step 2: determine multiple BP according to knearest neighbour method cluster Member's collection of neutral net；Step 3: determine the tuple of multi-BP Neural Networks according to Validity Index；Step 4: according to The analysis result of step one and step 2, three determine that the BP neutral net after tuple carries out load prediction.

Further, also include step 5: BP neutral net is improved, i.e. introduce factor of momentum, use and repeatedly calculate BP neutral net is improved by the mode averaged.

Further, BP neutral net is transmitted by the forward of signal and back propagation two parts of error form, and calculates reality output Time carry out by the direction being input to output, each layer weights, the makeover process of threshold value are then carried out from exporting to the direction inputted.

Further, the back propagation of described error includes: successively calculate the output error of each layer neuron by output layer, Regulate weights and the threshold value of each layer further according to error gradient descent method, after making regulation, final output of network mapping can be close to expected value.

Further, described step one specifically includes:

Tectonic sequence matrix, based on historical load data longitudinally 24 point load sequences, sets up initial load sequence matrix L=[L₁,L₂... Lm], wherein m numerical value is 24, and N is longitudinal dimension of historical load data, corresponding load record natural law,

Nondimensionalization, carries out data process by initial value method method, obtains dimensionless matrix, be denoted as L '=[L '₁,L′₂,…,L′_m], L′_i(k)=L_i(k)/L_i(1) i=1,2 ..., N；K=1,2 ..., m；

Coefficient of association calculates,In formula, p and q is longitudinal 24 The sequence number of point load sequence,For resolution ratio, k is longitudinal length index, ξ_pqK () is row k p row load and q row load Coefficient of association；

Calculation of relationship degree, γ_pqFor the degree of association of pth row load Yu q row load, by the association of the calculated load sequence of the degree of association Coefficient matrix,

Further, described step 2 specifically includes:

Euclidean distance quantitative scoring is used to calculate the distance vector characterizing load incidence matrix sequence similarity each other；

Knearest neighbour method is used to obtain the matrix comprising clustering tree information；

ConvolutionDetermined by the incidence coefficient matrix of load sequence, use short distance The pedigree diagram of multi-BP Neural Networks tuple is determined from method.

Further, the Validity Index in described step 3 isWherein, Lp (i), Lq (i) represent the load in i-th day p, q moment in same class respectively, Lr (i), Lt (i) represent respectively in inhomogeneity i-th day r, The load of t, N is longitudinal dimension of historical load data, corresponding load record natural law.

Compared with prior art, the invention has the beneficial effects as follows: propose based on grey relational grade and knearest neighbour method cluster multiple The method that the tuple of BP neutral net is selected, to merge part closely-related load sequence, suitably reduces multiple BP nerve net The tuple of network.Meanwhile, introduce factor of momentum, and employing repeatedly calculates the mode averaged, and improves BP neutral net and easily falls into The problem entering local convergence, improves its resistant and oscillation resistant ability.Multi-BP Neural Networks compares traditional BP neural network prediction model, There is more preferable prediction effect.

Accompanying drawing explanation

Fig. 1 is the three layers of BP neural network structure schematic diagram of typical case that the present invention relates to.

Fig. 2 is that pedigree diagram selected by member's collection of multi-BP Neural Networks in the present invention.

Fig. 3 is multi-BP Neural Networks model Short Term Load effect in the present invention.

Fig. 4 is the load prediction results figure of 6 weight BP neutral nets in the present invention.

Detailed description of the invention

The present invention is further detailed explanation with detailed description of the invention below in conjunction with the accompanying drawings.

1, traditional BP neural network prediction principle resolves

1) BP neutral net basic model

In 1986, the scientist headed by Rumelhart and McCelland proposed BP neutral net, and it is that one can be learned Practise and store substantial amounts of input-output mode map relation, and without disclosing the multilamellar of the math equation of this mapping relations in advance before Feedback neutral net, is made up of input layer, hidden layer and output layer.Fig. 1 is the structure of typical three layers of BP neutral net Figure, uses totally interconnected mode between layers, does not exist and be connected with each other between same layer, and hidden layer can be with one or more layers.Figure In 1, x_jRepresent the input of input layer jth node；w_ijRepresent that hidden layer i-th node is between input layer jth node Weights；θ_iThreshold value for hidden layer i-th node；φ is the excitation function of hidden layer；w_kiRepresent that output layer kth node arrives Weights between hidden layer i-th node；α_kThreshold value for output layer kth node；ψ is the excitation function of output layer；o_k Represent the output of kth node.

2) transmission of BP neutral net signal and error correction

Basic BP neural network algorithm is transmitted by the forward of signal and back propagation two parts of error form, and i.e. calculates actual defeated Carrying out by the direction being input to output when going out, each layer weights, the makeover process of threshold value are then carried out from exporting to the direction inputted. According to parameter shown in Fig. 1, the output signal of BP neutral net, each layer weights and threshold value are calculated and adjust.

(1) the propagated forward process of input signal

According to the structure chart of BP neutral net in Fig. 1, the input net of hidden layer i-th node of interest can be learnt_iWith output Amount o_i, input quantity net of output layer kth node_kWith output o_kIt is respectively

${\mathrm{net}}_i=\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{w}_{ij}{x}_j+{\mathrm{\θ}}_i---\left(1\right)$

$o_i=\mathrm{\φ}\left({\mathrm{net}}_i\right)=\mathrm{\φ}(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{ij}{x}_j+{\mathrm{\θ}}_j)---\left(2\right)$

${\mathrm{net}}_k=\underset{i=1}{\overset{q}{\mathrm{\Σ}}}{w}_{ki}\mathrm{\φ}(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{ij}{x}_j+{\mathrm{\θ}}_j)+{\mathrm{\α}}_k---\left(3\right)$

$o_k=\mathrm{\ψ}\left({\mathrm{net}}_k\right)=\mathrm{\ψ}\[\underset{i=1}{\overset{q}{\mathrm{\Σ}}}{w}_{ki}\mathrm{\φ}(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{ij}{x}_j+{\mathrm{\θ}}_j)+{\mathrm{\α}}_k\]---\left(4\right)$

(2) back-propagation process of error signal

The back propagation of error, first successively calculates the output error of each layer neuron, then according to error by output layer Gradient descent method regulates weights and the threshold value of each layer, and after making regulation, final output of network mapping can be close to expected value.According to error Gradient descent method, can revise hidden layer successively to output layer modified weight amount Δ w_ki, output layer threshold value correction amount α_k, input layer To hidden layer modified weight amount Δ w_ijWith hidden layer threshold value correction amount θ_iAs shown in formula (5)-formula (8), in formula, η is learning rate, P is training sample sum.

${\mathrm{\Δw}}_{ki}=\mathrm{\η}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{L}{\mathrm{\Σ}}}(T_k^p-o_k^p){\mathrm{\ψ}}^{\′}\left({\mathrm{net}}_k\right)o_i---\left(5\right)$

${\mathrm{\Δ\α}}_k=\mathrm{\η}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{L}{\mathrm{\Σ}}}(T_k^p-o_k^p){\mathrm{\ψ}}^{\′}\left({\mathrm{net}}_k\right)---\left(6\right)$

${\mathrm{\Δw}}_{ij}=\mathrm{\η}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{L}{\mathrm{\Σ}}}(T_k^p-o_k^p){\mathrm{\ψ}}^{\′}\left({\mathrm{net}}_k\right)w_{ki}{\mathrm{\φ}}^{\′}\left({\mathrm{net}}_i\right)x_j---\left(7\right)$

${\mathrm{\Δ\θ}}_i=\mathrm{\η}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{L}{\mathrm{\Σ}}}(T_k^p-o_k^p){\mathrm{\ψ}}^{\′}\left({\mathrm{net}}_k\right)w_{ki}{\mathrm{\φ}}^{\′}\left({\mathrm{net}}_i\right)---\left(8\right)$

2, load serial correlation based on grey relational grade is analyzed

Association analysis is that a kind of of gray system theory proposition analyzes the method for each correlate degree in system, and its basic thought is Judging correlation degree according to similarity degree between curve, calculation procedure is as follows:

1) tectonic sequence matrix.Based on historical load data longitudinally 24 point load sequences, set up initial load sequence matrix L=[L₁,L₂... Lm], wherein m numerical value is 24, and N is longitudinal dimension of historical load data, corresponding load record natural law.

${\left[\begin{array}{c}{L}_1\\ L_2\\ .\\ .\\ .\\ L_N\end{array}\right]}^T=\left[\begin{array}{ccc}{L}_1\left(1\right)& ...& L_1\left(m\right)\\ .& .& .\\ .& .& .\\ .& .& .\\ L_N\left(1\right)& ...& L_N\left(m\right)\end{array}\right]---\left(9\right)$

2) nondimensionalization.For eliminating the impact of dimension, carry out data process by initial value method method.Employing formula (10) can obtain dimensionless Matrix, is denoted as L '=[L '₁,L′₂,…,L′_m]。

L′_i(k)=L_i(k)/L_i(1) i=1,2 ..., N；K=1,2 ..., m (10)

3) coefficient of association calculates.

In formula: p and q is the sequence number of longitudinally 24 point load sequences；For resolution ratio, its role is to improve coefficient of association Between the significance of difference, typically taking its value is 0.5；K is longitudinal length index；ξ_pqK () is row k p row load and q row load Coefficient of association.

4) calculation of relationship degree.γ in formula (12)_pqThe degree of association for pth row load Yu q row load.By the calculated load of the degree of association The incidence coefficient matrix of sequence.

${\mathrm{\γ}}_{pq}=\frac{1}{N-1}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\ξ}}_{pq}\left(k\right),k=1,2,...,N---\left(12\right)$

3, member's collection of multi-BP Neural Networks is determined based on knearest neighbour method cluster

Cluster algorithm is numerous, such as K-Means, clustering algorithm based on grid and density, F clustering algorithm and beeline Method etc..Wherein, use knearest neighbour method to cluster, compare other algorithm simple and easy to operate.Therefore, knearest neighbour method is selected Carry out load Sequence clustering, and select suitable clustering distance amount, to determine the tuple of final multi-BP Neural Networks.Entering During row cluster, Euclidean distance quantitative scoring is first used to calculate the distance vector characterizing load incidence matrix sequence similarity each other, Secondly knearest neighbour method is used to obtain the matrix comprising clustering tree information.The coefficient of association of load sequence determined by convolution (12) Matrix, uses knearest neighbour method to can determine that the pedigree diagram of multi-BP Neural Networks tuple.

4, the tuple of multi-BP Neural Networks is determined based on Validity Index

For weighing the quality of cluster result, characterize, from cluster Essential, the Validity Index that Clustering Effect is good and bad.In view of more excellent Cluster result should have the feature that the spacing of class inside is the smaller the better, between class distance is the bigger the better, definition is as shown in formula (13) Validity Index.

In formula: L_p(i)、L_qI () represents the load in i-th day p, q moment in same class respectively；L_r(i)、L_tI () represents inhomogeneity respectively In i-th day r, the load of t；The same formula of N implication (9).

Below by instantiation, the inventive method and beneficial effect are described in more detail.

1, example system and data process

Load data that Data Source of the present invention is gathered in certain actual electric network and weather data, the sampling time interval of each equipment Cycle is 1h, and Weather information is dry-bulb temperature, dew point temperature.Although laboratory data amount is not reaching to the scale of big data, but Correctness of algorithm experiment can be carried out by this experimental data, provide a kind of new method for the load prediction under big data environment.Instruction Practice the electricity consumption data that scope is on March 31st, 1 day 1 January in 2014.Prediction day is that on April 1st, 2014 is different The electric load in moment, as shown in table 1.The research of load data is found these data present a kind of continuity, periodically, The feature of dependency, according to the achievement in research of these features and lot of documents determine sample attribute be day the last fortnight load in the same time, Day the last week load in the same time, day load the most in the same time, the previous day day of load in the same time, the previous day day of dry bulb temperature in the same time Degree, the previous day day of dew point temperature in the same time, dry-bulb temperature in the same time on the prediction same day and dew point temperature in the same time on the prediction same day, it was predicted that Day actual load in the same time, its sample data is as shown in table 2.

The actual load data in table on April 1st, 1 2014

Moment/h	Load/MW	Moment/h	Load/MW	Moment/h	Load/MW
						1	11483	9	15870	17	14508
2	10924	10	15965	18	14332
						3	10711	11	15978	19	14219
4	10728	12	15823	20	14702
						5	11027	13	15556	21	15265
6	12128	14	15388	22	14557
						7	14043	15	15060	23	13416
8	15413	16	14761	24	12135

Table 2 weight training set of data samples

Attribute	Value
		Day the last fortnight load in the same time	11600MW
Day the last week load in the same time	11857MW
		Day load the most in the same time	11462MW
Day the previous day load in the same time	11203MW
		Day the previous day dry-bulb temperature in the same time	46℃
Day the previous day dew point temperature in the same time	43℃
		Predict dry-bulb temperature in the same time on the same day	41℃
Predict dew point temperature in the same time on the same day	18℃
		Predict day actual load in the same time	11483MW

It should be noted that BP neutral net logarithm value number between 0 with 1 compares sensitive, therefore, by original loads Before sequence inputting distributed BP neural network model, needing first to be normalized data, training is returned counter after terminating One change processes, and obtains actual load predictive value.

2, experimental result and analysis

This experiment purpose is to confirm that multi-BP Neural Networks model is compared one BP of all historical load data sharings directly Neural network model, has more preferable precision of prediction.Convolution (13) learns the load prediction mould of the BP neutral net setting up 6 weights Type is relatively reasonable, and corresponding Validity Index value is minimum, is 0.1422.Tuple member selection, according to can refer to Fig. 2, i.e. sets When being set to 6, it is meant that create 6 BP load forecasting model, by numbered in numbered for moment 2-8, moment 9, moment numbering It is that the load correlated series in 1 numbered with 20-24, moment 10, moment numbered 11 is respectively as BP load forecasting model 1-5 Input, output, the load correlated series in residue moment is as the input of BP load forecasting model 6, output.

Fig. 3 prediction effect figure is that the result averaged is run multiple times.Can learn that BP neutral net tuple is to load prediction precision Impact relatively big, set up the Short Term Load models of 6 weight BP neutral nets, compare all historical load correlated serieses and share Same BP neural network prediction model, has more preferable precision of prediction.Load prediction effects based on 6 weight BP neutral nets are such as Shown in Fig. 4, mean absolute relative error is 3.01%, and root-mean-square error is 1.63%, meets wanting of actual load precision of prediction Ask.

Claims (7)

1. a multi-BP Neural Networks load forecasting method based on grey relational grade, it is characterised in that comprise the following steps: step Rapid one: use grey relevant degree method that the relatedness of load sequence is analyzed；Step 2: according to knearest neighbour method cluster really Determine member's collection of multi-BP Neural Networks；Step 3: determine the tuple of multi-BP Neural Networks according to Validity Index；Step Four: according to the analysis result of step one and step 2, three determine that the BP neutral net after tuple carries out load prediction.

2. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 1, it is characterised in that also Including step 5: BP neutral net is improved, i.e. introduce factor of momentum, use and repeatedly calculate the mode pair averaged BP neutral net improves.

3. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 1 or 2, it is characterised in that Described BP neutral net is transmitted by the forward of signal and back propagation two parts of error form, and calculates when reality exports by input Direction to output is carried out, and each layer weights, the makeover process of threshold value are then carried out from exporting to the direction inputted.

4. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 3, it is characterised in that institute The back propagation stating error includes: successively calculate the output error of each layer neuron by output layer, further according under error gradient Fall method regulates weights and the threshold value of each layer, and after making regulation, final output of network mapping can be close to expected value.

5. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 1 or 2, it is characterised in that Described step one specifically includes:

Tectonic sequence matrix, based on historical load data longitudinally 24 point load sequences, sets up initial load sequence matrix L=[L₁,L₂... Lm], wherein m numerical value is 24, and N is longitudinal dimension of historical load data, corresponding load record natural law,

Nondimensionalization, carries out data process by initial value method method, obtains dimensionless matrix, be denoted as L '=[L '₁,L′₂,…,L′_m], L′_i(k)=L_i(k)/L_i(1) i=1,2 ..., N；K=1,2, m；

Coefficient of association calculates,In formula, p and q is longitudinal 24 The sequence number of point load sequence, θ is resolution ratio, and k is longitudinal length index, ξ_pqK () is row k p row load and q row load Coefficient of association；

Calculation of relationship degree, γ_pqFor the degree of association of pth row load Yu q row load, by the association of the calculated load sequence of the degree of association Coefficient matrix,

6. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 1 or 2, it is characterised in that Described step 2 specifically includes:

Euclidean distance quantitative scoring is used to calculate the distance vector characterizing load incidence matrix sequence similarity each other；

Knearest neighbour method is used to obtain the matrix comprising clustering tree information；

ConvolutionDetermined by the incidence coefficient matrix of load sequence, use short distance The pedigree diagram of multi-BP Neural Networks tuple is determined from method.

7. multi-BP Neural Networks load forecasting method based on grey relational grade as claimed in claim 1 or 2, it is characterised in that Validity Index in described step 3 isWherein, Lp (i), Lq (i) Representing the load in i-th day p, q moment in same class respectively, Lr (i), Lt (i) represent in inhomogeneity i-th day r, t respectively Load, N is longitudinal dimension of historical load data, corresponding load record natural law.