CN106022549A - Short term load predication method based on neural network and thinking evolutionary search - Google Patents

Abstract

The invention provides a short term load predication method based on a neural network and thinking evolutionary search. The method comprises the steps of (1) obtaining historical load data, and carrying out segmentation and normalization on the data, (2) obtaining weather conditions, data types and other related data samples which influence the load predication, and taking the data samples and the historical load data as input variables, (3) determining the input and output samples of a training set and a testing set, (4) according to the matrix parameter of the input and output samples, determining Elman network input and output layer neuron numbers, (5) determining an Elman network structure, (6) determining an encoding length by a network structure, (7) operating MEA algorithm parameter setting, (8) carrying out assimilation and dissimilation operations to obtain an optimal weight and an optimal threshold, (9) establishing an Elman network model by the optimal weight and the optimal threshold, and carrying out load predication, and (10) carrying out normalization process on the network output value, obtaining a predication value, and carrying out error analyzing and predication performance evaluation.

Description

Based on neutral net and the short-term load forecasting method of thinking evolution search

Technical field

The invention belongs to system for distribution network of power load prediction technical field, be specifically related to a kind of based on neutral net and think of The short-term load forecasting method of dimension evolution search.

Background technology

Load forecast is the important evidence of the departments such as power planning, marketing, market transaction, scheduling, preferably Forecasting Methodology contributes to power department and makes correct decision-making.According to predicted time length, generally can be by power system load Prediction is divided into the several types such as long-term, mid-term, short-term and ultra-short term prediction.

Wherein short-term load forecasting is the important component part of load forecast, mainly includes that daily load prediction and week are born Lotus is predicted, is respectively used to arrange daily dispatch scheduling and week operation plan.Under current Electricity Market, improve short term Precision of prediction, to arranging startup-shutdown plan, unit commitment, generating capacity to make rational planning for and economic load dispatching and electric power city Field transaction etc. has important practical significance.

Experts and scholars propose multiple Forecasting Methodology the most both at home and abroad, mainly may be used according to the difference of prediction Mathematical Modeling Methods To be divided into two big classes: mathematical statistics method and artificial intelligence method.Mathematical statistics method specifically include that random sequence method, linear regression method, State space method and exponential smoothing etc..Mathematical statistics method can preferably carry out linear load prediction, and lacks process The ability of nonlinear-load.Artificial intelligence approach includes: expert system approach, fuzzy reasoning method, artificial neural network method etc..Expert Systems approach need to obtain experiential operating knowledge, and it is relatively difficult that Heuristics is converted into mathematical programming.Fuzzy reasoning is special The extension of family's system, in order to obtain the non-linear behavior of load data, needs to be tied by expertise structure fuzzy reasoning optimum Structure and minimize error model.Artificial neural network is independent of artificial experience, by training sample data, sets up prediction network, Can preferably process the non-linear relation of load data.

Through retrieving, Load Prediction In Power Systems method based on BP neutral net (Publication No. CN 103295081 A), And a kind of resident load Forecasting Methodology based on Elman neutral net of patent (Publication No. CN 104636822 A), exist Following problem: owing to BP neutral net and Elman network all use BP algorithm, above patent does not considers that BP algorithm is in study During new samples, there is the trend forgeing old sample, and when asking for weight threshold, there is the defect being easily trapped into local minimizers number, Affect load prediction precision.

The present invention have selected Elman artificial neural network, has impermanent memory function, it is possible to avoids network training new sample In this time, forgets the drawback of old sample.Utilize global optimizing and the parallel search performance of mind evolutionary (MEA) simultaneously, optimize The weights of Elman neutral net and threshold value, establish MEA-Elman forecast model, improve predictablity rate.

Summary of the invention

It is an object of the invention to provide the short-term load forecasting method of a kind of search of developing based on neutral net and thinking.

The present invention is based on neutral net and the short-term load forecasting method of thinking evolution search, the steps include:

Step one: obtain historical load number data, and it is carried out segmentation and normalized；

Step 2: obtain the data sample of the relative influence load prediction such as weather conditions and date type, together with historical load number According to together as input variable；

Step 3: determine the input and output sample of training set, test set；

Step 4: according to the matrix dimension of input and output sample, determines Elman network input layer neuron number S₁And output layer Neuron number S₃, middle hidden layer neuron number is determined by network repetition training, is designated as S₂；

Step 5: determine Elman network structure, Elman network is made up of, wherein input layer, hidden layer, undertaking layer, output layer Accept layer and be used for remembering the output of hidden layer previous moment, and this output information is fed back to hidden layer, generally can be by this network Structure is abbreviated as S₁-S₂-S₃；

Step 6: determined code length by network structure, is designated as S, the i.e. required weight threshold number optimized；

Step 7: carry out MEA algorithm parameter setting, Population Size POP is set_size, winning population number Best_size, plant temporarily Group's size Tem_size, sub-population size S_G；

Step 8: carry out convergent and operation dissimilation, compares its fitness value Fitness size and produces winning individuality, obtains Excellent weight w₁、w₂、w₃, optimal threshold b₁、b₂；

Step 9: by best initial weights and threshold value w₁、w₂、w₃、b₁、b₂Set up Elman network model, and carry out load prediction；

Step 10: network output valve is gone normalized, obtains predictive value, and carries out error analysis and estimated performance is commented Estimate.

The forecast model of the present invention provides the benefit that: the present invention utilizes MEA algorithm optimization Elman network, it is to avoid network exists Ask for weight threshold and be easily absorbed in the defect of local minimizers number, and then improve the precision of load prediction, to power system rational management, Market planning and raising power consumption efficiency, important in inhibiting.

Accompanying drawing explanation

Fig. 1 is forecast model structured flowchart of the present invention, and Fig. 2 is the MEA algorithm flow chart that the present invention uses, and Fig. 3 is this The Elman neural network structure figure of bright use, Fig. 4 is the MEA-Elman Optimized model flow chart that the present invention uses, and Fig. 5 is this The MEA-Elman forecast model that invention proposes predicts the outcome and Elman neural network prediction Comparative result figure, and Fig. 6 is the present invention The MEA-Elman forecast model proposed predicts the outcome and BP neural network prediction Comparative result figure, and Fig. 7 is the electric power of load prediction Big data platform Organization Chart, Fig. 8 is MapReduce parallel computation schematic diagram, and Fig. 9 is the MEA-Elman prediction that the present invention proposes The prediction block diagram that algorithm is combined with MapReduce.

Detailed description of the invention

Embodiment one, the present invention is based on neutral net and the short-term load forecasting method of thinking evolution search, its step For:

Step one: obtain historical load number data, and it is carried out segmentation and normalized；

Step 2: obtain the data sample of the relative influence load prediction such as weather conditions and date type, together with historical load number According to together as input variable；

Step 3: determine the input and output sample of training set, test set；

Step 4: according to the matrix dimension of input and output sample, determines Elman network input layer neuron number S₁And output layer Neuron number S₃, middle hidden layer neuron number is determined by network repetition training, is designated as S₂；

Step 5: determine Elman network structure, Elman network is made up of, wherein input layer, hidden layer, undertaking layer, output layer Accept layer and be used for remembering the output of hidden layer previous moment, and this output information is fed back to hidden layer, generally can be by this network Structure is abbreviated as S₁-S₂-S₃；

Step 6: determined code length by network structure, is designated as S, the i.e. required weight threshold number optimized；

Step 7: carry out MEA algorithm parameter setting, Population Size POP is set_size, winning population number Best_size, plant temporarily Group's size Tem_size, sub-population size S_G；

Step 8: carry out convergent and operation dissimilation, compares its fitness value Fitness size and produces winning individuality, obtains Excellent weight w₁、w₂、w₃, optimal threshold b₁、b₂；

Step 9: by best initial weights and threshold value w₁、w₂、w₃、b₁、b₂Set up Elman network model, and carry out load prediction；

Step 10: network output valve is gone normalized, obtains predictive value, and carries out error analysis and estimated performance is commented Estimate.

The above Forecasting Methodology, in step one, historical load data sectional rule is: by the value in L moment in the past, enter Row P walks prediction, and desirable L adjacent sample is sliding window, and is mapped to P predictive value, i.e. input is: { [X₁、 X₂……X_L]、[X₂、X₃……X_L+1]…[X_K、X_K+1……X_K+L-1], then correspondence is output as: { [X_L+1、X_L+2……X_L+P]、 [X_L+2、X_L+3……X_L+P+1]…[X_K+L、X_K+L+1……X_K+L+P-1]}；Here L=3, P=1 are taken, i.e. with the load data of first three day Predict the 4th day load variations situation.

The above Forecasting Methodology, in step one, normalized expression formula is:

$y=\frac{(y_{max}-y_{min})*(x-x_{min})}{x_{\max }-x_{\min }}+y_{min}---\left(1\right)$

Wherein y_minAnd y_maxIt is normalized parameter, takes y_min=0, y_max=1；Y is standard value after normalization, and span is [y_min, y_max]；And x is sample original value, x_min、x_maxIt is respectively the minima in sample and maximum.

The above Forecasting Methodology, in step 6, code length S expression formula is:

S=S₁*S₂+S₂*S₂+S₂*S₃+S₂+S₃ (2)

Wherein S₁*S₂For the number connecting weights of input layer to hidden layer, S₂*S₂For accepting the layer connection weights to hidden layer Number, S₂*S₃For the connection weights number of hidden layer to output layer, S₂For hidden layer threshold number, S₃For output layer threshold value Number.

The above Forecasting Methodology, sub-population size S in step 7_GExpression formula is:

S_G=POP_size/(Best_size+Tem_size) (3)

Wherein POP_sizeFor Population Size, Best_sizeWinning population number, Tem_sizeInterim population number.

The above Forecasting Methodology, in step 8, fitness value function Fitness expression formula is:

$Fitness=1/\frac{{(T_{sim}-T_{test})}^2}{n}---\left(4\right)$

Wherein T_testFor test set actual value, T_simFor test set predictive value, n is predictive value number.

The above Forecasting Methodology, goes the normalized expression formula to be in step 10:

$x_F=\frac{(x_{Tmax}-x_{Tmin})*(y_o-y_{min})}{y_{max}-y_{\min }}+x_{Tmin}---\left(5\right)$

Wherein x_FFor removing the return value after normalization, x_TmaxAnd x_TminIt is respectively the maximum in test set sample and minima, y_o For network output valve, parameter y_minAnd y_maxValue keep constant, be taken as 0 and 1 respectively.

The above Forecasting Methodology, in step 10, error analysis expression formula is:

$MAPE=\frac{1}{N}\left(\underset{i=1}{\overset{N}{\Σ}}\right|\frac{{\mathrm{Re}}_i-F_i}{{\mathrm{Re}}_i}|*100\%)---\left(6\right)$

$RMSE=\sqrt{\frac{\underset{i=1}{\overset{N}{\Σ}}{(F_i-{\mathrm{Re}}_i)}^2}{N}}---\left(7\right)$

Wherein Re_iFor actual value, F_iFor predictive value, N is predictive value number；MAPE is mean percent absolute error (Mean Absolute Percentage Error)；RMSE is root-mean-square error (Root Mean Square Error).

Below the method for the present invention is further launched, concretely comprising the following steps of the method for the present invention:

(1) obtain historical load number data, and it is carried out segmentation and normalized；

(2) data sample of the relative influence load prediction such as weather conditions and date type is obtained, together with historical load data one Rise as input variable；

(3) the input and output sample of training set, test set is determined；

(4) according to the matrix dimension of input and output sample, Elman network input layer neuron number S is determined₁Neural with output layer Unit's number S₃, middle hidden layer neuron number is determined by network repetition training, is designated as S₂；

(5) determining Elman network structure, Elman network is made up of input layer, hidden layer, undertaking layer, output layer, wherein accepts Layer is used for remembering the output of hidden layer previous moment, and this output information is fed back to hidden layer, generally can be by this network structure It is abbreviated as S₁-S₂-S₃；

(6) determined code length by network structure, be designated as S, the i.e. required weight threshold number optimized；

(7) carry out MEA algorithm parameter setting, Population Size POP is set_size, winning population number Best_size, interim population big Little Tem_size, sub-population size S_G；

(8) carry out convergent and operation dissimilation, compare its fitness value Fitness size and produce winning individuality, obtain optimum power Value w₁、w₂、w₃, optimal threshold b₁、b₂；

(9) by best initial weights and threshold value w₁、w₂、w₃、b₁、b₂Set up Elman network model, and carry out load prediction；

(10) network output valve is gone normalized, obtain predictive value, and carry out error analysis and forecast performance evaluation.

Wherein in step (1), historical load data sectional rule is: by the value in L moment in the past, carry out P step prediction, can Taking L adjacent sample is sliding window, and is mapped to P predictive value, i.e. input is: { [X₁、X₂……X_L]、[X₂、 X₃……X_L+1]…[X_K、X_K+1……X_K+L-1], then correspondence is output as: { [X_L+1、X_L+2……X_L+P]、[X_L+2、X_L+3…… X_L+P+1]…[X_K+L、X_K+L+1……X_K+L+P-1]}；Here take L=3, P=1, i.e. predict the 4th day with the load data of first three day and bear Lotus situation of change.

Wherein in step (1), the expression formula of normalized is as follows:

$y=\frac{(y_{max}-y_{min})*(x-x_{min})}{x_{max}-x_{min}}+y_{min}---\left(1\right)$

Wherein y_minAnd y_maxIt is normalized parameter, takes y_min=0, y_max=1；Y is standard value after normalization, and span is [y_min, y_max]；And x is sample original value, x_min、x_maxIt is respectively the minima in sample and maximum.

Wherein in step (6), code length S expression formula is as follows:

S=S₁*S₂+S₂*S₂+S₂*S₃+S₂+S₃ (2)

In formula, S₁*S₂For the number connecting weights of input layer to hidden layer, S₂*S₂For accepting the layer connection weights to hidden layer Number, S₂*S₃For the connection weights number of hidden layer to output layer, S₂For hidden layer threshold number, S₃For output layer threshold value Number.

Wherein step (7) sub-population size S_GExpression formula is:

S_G=POP_size/(Best_size+Tem_size) (3)

In formula, POP_sizeFor Population Size, Best_sizeWinning population number, Tem_sizeInterim population number.

Wherein step (8) fitness value function Fitness expression formula is:

$Fitness=1/\frac{{(T_{sim}-T_{test})}^2}{n}---\left(4\right)$

Wherein T_testFor test set actual value, T_simFor test set predictive value, n is predictive value number.

Wherein step (9) is with w₁、w₂、w₃、b₁、b₂Assignment expression for weight threshold is:

Net_optimized.IW{1,1}=w₁ (5)

Net_optimized.LW{1,1}=w₂ (6)

Net_optimized.LW{2,1}=w₃ (7)

Net_optimized.b{1}=b₁ (8)

Net_optimized.b{2}=b₂ (9)

(5) formula represents w₁It is assigned to the input layer connection weights to implicit interlayer；(6) formula represents w₂It is assigned to accept layer arrive The connection weights of implicit interlayer；(7) formula represents w₃It is assigned to the hidden layer connection weights to output interlayer；(8) formula represents b₁ It is assigned to the threshold value of hidden layer；(9) formula represents b₂It is assigned to the threshold value of output layer；(5) net_optimized～in (9) formula Represent the weight threshold after MEA algorithm optimization.

Wherein in step (9), Elman network model state-space expression is as follows:

y_h(k)=g (w₁u(k-1)+w₂y_c(k)+b₁) (10)

y_c(k)=y_h(k-1) (11)

y_o(k)=f (w₃y_h(k)+b₂) (12)

(10) y in formula_hK () is hidden layer output, w₁Connecting weights for input layer to hidden layer, u (k-1) is input sample, w₂For Accept the layer connection weights to implicit interlayer, y_cK () is for accepting layer output, b₁For hidden layer threshold value, g (k) is that hidden layer transmits letter Number, can be chosen as Sigmoid function, and expression formula is: g (k)=1/ (1+e^-k)；(11) y in formula_cK () is for accepting layer output, y_h (k-1) it is the output of hidden layer previous moment；(12) y in formula₀K () is output layer output, w₃Connection weight for hidden layer to output layer Value, y_hK () is hidden layer output, b₂For output layer threshold value, f (k) is that output layer transmits function, generally linear function.

Wherein step (10) is gone the normalized expression formula to be:

$x_F=\frac{(x_{Tmax}-x_{Tmin})*(y_o-y_{min})}{y_{max}-y_{\min }}+x_{Tmin}---\left(13\right)$

Wherein x_FFor removing the return value after normalization, x_TmaxAnd x_TminIt is respectively the maximum in test set sample and minima, y_o For network output valve, parameter y_minAnd y_maxValue keep constant, be taken as 0 and 1 respectively.

Wherein in step (10), error expression is as follows:

$MAPE=\frac{1}{N}\left(\underset{i=1}{\overset{N}{\Σ}}\right|\frac{\mathrm{Re}-F_i}{{\mathrm{Re}}_i}|*100\%)---\left(14\right)$

$RMSE=\sqrt{\frac{\underset{i=1}{\overset{N}{\Σ}}{(F_i-{\mathrm{Re}}_i)}^2}{N}}---\left(15\right)$

(14), Re in (15) formula_iFor actual value, F_iFor predictive value, N is predictive value number；(14) in formula, MAPE is mean percent Absolute error (MeanAbsolute Percentage Error)；(15) in formula, RMSE is root-mean-square error (Root Mean Square Error)；MAPE, RMSE can be simultaneously as the comprehensive value model of prediction；MAPE, RMSE value are the least, show prediction Effect is the best.

Describe specific embodiments of the present invention below in conjunction with the accompanying drawings in detail.

This embodiment uses somewhere integral point load every day first 30 day of January, weather conditions and date type thereof for number According to sample, it was predicted that the 31st day integral point moment load.

Fig. 1 is forecast model structured flowchart of the present invention, and the present invention is by weather conditions, date type and historical load number According to the input as MEA-Elman forecast model, carry out load prediction.Wherein weather conditions can represent with normalized parameter, as Shown in table 1.Date type normalized parameter is as shown in table 2.

Table 1 weather conditions normalized parameter:

Weather conditions	Normalized parameter	Weather conditions	Normalized parameter
				Fine	0	Drizzle or moderate rain	0.4
Clear to cloudy	0.1	Moderate rain	0.5
				Cloudy	0.2	Moderate rain is to heavy rain	0.6
Cloudy	0.25	Heavy rain	0.7
				Cloudy turn to overcast	0.25	Heavy rain	0.8
Light rain	0.3	Extra torrential rain	0.9

Table 2 date type normalized parameter:

Date type	Normalized parameter	Date type	Normalized parameter
				Monday	0.1	Friday	0.5
Tuesday	0.2	Saturday	0.6
				Wednesday	0.3	Sunday	0.7
Thursday	0.4	Festivals or holidays	0.6

Take the load data of 24 every day, produce the data sample of 31 × 24, be designated as:

${\mathrm{Data}}_{31\×24}=\left[\begin{array}{cccc}{D}_{1\×1}& D_{1\×2}& ...& D_{1\×24}\\ D_{2\×1}& D_{2\×2}& ...& D_{2\×24}\\ ...& ...& ...& ...\\ D_{31\×1}& D_{31\×2}& ...& D_{31\×24}\end{array}\right]---\left(16\right)$

Data sample is carried out segment processing, predicts the integral point load data of the 4th day with the integral point load data of first 3 days, now Form new data sample, be designated as:

$P_{28\×63}=\left[\begin{array}{ccccccccc}{D}_{1\×1}& \mathrm{..}& D_{1\×24}& D_{2\×1}& \mathrm{..}& D_{2\×24}& D_{3\×1}& \mathrm{..}& D_{3\×24}\\ D_{2\×1}& \mathrm{..}& D_{2\×24}& D_{3\×1}& \mathrm{..}& D_{3\×24}& D_{4\×1}& \mathrm{..}& D_{4\×24}\\ \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}& \mathrm{..}\\ D_{28\×1}& \mathrm{..}& D_{28\×24}& D_{29\×1}& \mathrm{..}& D_{29\×24}& D_{30\×1}& \mathrm{..}& D_{30\×24}\end{array}\right]---\left(17\right)$

Using 1 to 27 row of P matrix as the input of training set, it is designated as P_train；Using Data matrix 4 to 30 row as training set Output, is designated as T_train；Using the 28th row of P matrix as the input of test set, it is designated as P_test；Using the 31st row of Data matrix as The output of test set, is designated as T_test.Consider weather conditions every day and date type thereof simultaneously, just can 30 days in the past load datas pre- Survey the load data in the 31st day 24 integral point moment.

Determining network each layer neuron number, this patent uses the 4th day load of integral point load prediction in first 3 days, and every day is little every 1 In time, be acquired once to load, within one day, amounts to 24 groups of data, within three days, amounts to 72 groups of data, it is considered to weather conditions and date type These two groups of data, then can determine that network input layer neuron number is 74, and network output layer neuron number is 24, hidden layer god Through unit, number is set to 97 through network through network repetition training.

Fig. 2 is mind evolutionary flow chart, first carries out MEA algorithm initialization, produces initial population, produces the most at random Raw winning sub-group and interim sub-group, then carry out convergent operation dissimilation, finally select the sub-group that fitness value is the highest, output Optimum individual.The most convergent is to carry out in sub-group, calculates all ideal adaptation angle value, and fitness value soprano is winning Body, remaining individuality learns to winner, and produces new sub-group, repeats above operation, until this sub-group is ripe；Alienation is Carry out in whole solution space, the fitness value of interim sub-group is compared with the fitness value of winning sub-group, if higher than winning Sub-group fitness value, then substitute this winning sub-group, otherwise, then this interim sub-group is abandoned.MEA algorithm and genetic algorithm Compare, the evolution information of a more than generation can be remembered, structure has parallel characteristics, calculate with convergent and alienation and instead of heredity Intersection in algorithm and mutation operation, improve efficiency of evolution.

Fig. 3 is Elman neural network structure figure, and Elman neutral net is a kind of dynamical feedback network, including input layer, Hidden layer, undertaking layer, output layer.Wherein accepting layer is a special hidden layer, and this layer receives feedback signal from hidden layer, logical Cross and coupled memory the hidden layer state in a upper moment is inputted the input together as hidden layer, quite together with the network of current time In feedback of status.It has impermanent memory function compared with traditional BP neural network prediction model, it is possible to avoid network in training The defect of old sample is forgotten during new samples.

Fig. 4 is MEA-Elman Optimized model flow chart, and specific works flow process is as follows:

Step1: produce training set/test set；

Step2: determine network structure, and MEA algorithm is carried out parameter setting；

Step3: produce initial population, winning sub-population and interim sub-population；

Convergent and the operation dissimilation of Step4: sub-population；

Step5: according to fitness value size, find out optimum individual；

Step6: set up Elman network by the weight threshold of optimum individual；

Step7: predict and emulate, analyses and prediction result, preserve optimum MEA-Elman model.

Fig. 5, Fig. 6 respectively present invention propose MEA-Elman forecast model predict the outcome with Elman neutral net and BP neural network prediction Comparative result figure.MEA-Elman model prediction accuracy is apparently higher than Elman network model as seen from the figure, And traditional BP neural network model.

Additionally, the predicting the outcome, as shown in table 3 of tri-kinds of different models of MEA-Elman, Elman, BP.

Table 3MEA-Elman, Elman, BP predict the outcome analysis:

By table 3, the MAPE value of each model, RMSE value can be calculated, as shown in table 4.

Table 4MEA-Elman, Elman, BP Model Error Analysis:

As shown in Table 4, the MEA-Elman model that the present invention provides has more preferable precision of prediction.It is compared with Elman network, Show the feasibility of this optimized algorithm；Compared with BP network, show this forecast model than traditional BP neural network model in advance Survey and had bigger improvement in precision.This forecast model improves load prediction precision, to realizing the generating rational management of capacity, warp Ji scheduling, power market transaction have practical significance.

Embodiment two, along with the fast development of digital information, quantity of information is explosive increase situation, and the whole world has been enter into greatly Data age.The big data of electric power are not only the deep application in power industry of the big data technique, are also power generation, consumption and phase Close the depth integration of technological revolution and big data theory, acceleration is advanced power industry development and business model innovation.

Fig. 7 is the electric power big data platform Organization Chart including load prediction, first pass through sensor, intelligent electric meter and The collecting units such as SCADA obtain the big data of electric power, then carry out cloud storage in distributed file system, then by cloud computing logarithm According to carrying out parallelization Treatment Analysis, finally the result of analysis is applied to decision-making, markets and the field such as scheduling.

Load prediction is one of important application technology of the big data of electric power, is basic needed for the marketing and power scheduling Information.This embodiment is from parallel load prediction, MEA-cloud computing programming model MapReduce and this patent proposed Elman prediction algorithm combines, and can improve MEA-Elman prediction algorithm and process ability and the precision of prediction of mass data.Cause MEA algorithm and neutral net have the natural talent to parallel data processing, therefore by MapReduce parallel computation frame and MEA- Elman prediction algorithm combines and has feasibility.

Fig. 8 is MapReduce parallel computation schematic diagram, be broadly divided into the Input stage, the Map stage, the Shuffle stage, Reduce stage and Output stage.The Input stage is to read data set from distributed file system, and is divided by this data set Sheet processes；The Map stage regards one group of (key, value) key-value pair, the Map letter then write by user as the data fragmentation of input Number runs, processes, and generates key-value pair in the middle of new (Key, value)；The Shuffle stage turns middle key-value pair from Map node Move on to the sequence of Reduce node and key assignments；The Reduce stage travels through all intermediate values, the Reduce function write by user Merge the value that same keys is corresponding；The Output stage i.e. exports corresponding result of calculation, and stores relevant position.

Fig. 9 is the prediction block diagram that the MEA-Elman prediction algorithm that the present invention proposes is combined with MapReduce.Magnanimity is born Lotus data set carries out piecemeal process, and is distributed to each MEA-Elman prediction module and is trained, i.e. the Map stage；Then will The training result in Map stage is transferred to Reduce node and merges process；Finally determine final forecast model, according to required Data sample carry out load prediction.

Claims (8)

1. based on neutral net and the short-term load forecasting method of thinking evolution search, it is characterised in that the steps include:

Step one: obtain historical load number data, and it is carried out segmentation and normalized；

Step 2: obtain the data sample of the relative influence load prediction such as weather conditions and date type, together with historical load number According to together as input variable；

Step 3: determine the input and output sample of training set, test set；

Step 4: according to the matrix dimension of input and output sample, determines Elman network input layer neuron number S₁And output layer Neuron number S₃, middle hidden layer neuron number is determined by network repetition training, is designated as S₂；

Step 5: determine Elman network structure, Elman network is made up of, wherein input layer, hidden layer, undertaking layer, output layer Accept layer and be used for remembering the output of hidden layer previous moment, and this output information is fed back to hidden layer, generally can be by this network Structure is abbreviated as S₁-S₂-S₃；

Step 6: determined code length by network structure, is designated as S, the i.e. required weight threshold number optimized；

Step 7: carry out MEA algorithm parameter setting, Population Size POP is set_size, winning population number Best_size, interim population Size Tem_size, sub-population size S_G；

Step 8: carry out convergent and operation dissimilation, compares its fitness value Fitness size and produces winning individuality, obtains Excellent weight w₁、w₂、w₃, optimal threshold b₁、b₂；

Step 9: by best initial weights and threshold value w₁、w₂、w₃、b₁、b₂Set up Elman network model, and carry out load prediction；

Step 10: network output valve is gone normalized, obtains predictive value, and carries out error analysis and estimated performance is commented Estimate.

2. Forecasting Methodology as claimed in claim 1, it is characterised in that step one historical load data sectional rule is: with L in the past The value in individual moment, carries out P step prediction, and desirable L adjacent sample is sliding window, and is mapped to P predictive value, the most defeated Enter for { [X₁、X₂……X_L]、[X₂、X₃……X_L+1]…[X_K、X_K+1……X_K+L-1], then correspondence is output as: { [X_L+1、 X_L+2……X_L+P]、[X_L+2、X_L+3……X_L+P+1]…[X_K+L、X_K+L+1……X_K+L+P-1]}；Here take L=3, P=1, i.e. use first three It load data predicts the 4th day load variations situation.

3. Forecasting Methodology as claimed in claim 1, it is characterised in that in step one, normalized expression formula is:

$y=\frac{(y_{max}-y_{min})*(x-x_{min})}{x_{\max }-x_{\min }}+y_{min}---\left(1\right)$

Wherein y_minAnd y_maxIt is normalized parameter, takes y_min=0, y_max=1；Y is standard value after normalization, and span is [y_min, y_max]；And x is sample original value, x_min、x_maxIt is respectively the minima in sample and maximum.

4. Forecasting Methodology as claimed in claim 1, it is characterised in that in step 6, code length S expression formula is:

S=S₁*S₂+S₂*S₂+S₂*S₃+S₂+S₃ (2)

Wherein S₁*S₂For the number connecting weights of input layer to hidden layer, S₂*S₂For accepting the layer connection weights to hidden layer Number, S₂*S₃For the connection weights number of hidden layer to output layer, S₂For hidden layer threshold number, S₃For output layer threshold number.

5. Forecasting Methodology as claimed in claim 1, it is characterised in that sub-population size S in step 7_GExpression formula is:

S_G=POP_size/(Best_size+Tem_size) (3)

Wherein POP_sizeFor Population Size, Best_sizeWinning population number, Tem_sizeInterim population number.

6. Forecasting Methodology as claimed in claim 1, it is characterised in that in step 8, fitness value function Fitness expression formula is:

$Fitness=1/\frac{{(T_{sim}-T_{test})}^2}{n}---\left(4\right)$

Wherein T_testTest set load actual value, T_simFor test set predictive value, n is predictive value number.

7. as claimed in claim 1 Forecasting Methodology, it is characterised in that go the normalized expression formula to be in step 10:

$x_F=\frac{(x_{Tmax}-x_{Tmin})*(y_o-y_{min})}{y_{max}-y_{\min }}+x_{Tmin}---\left(5\right)$

Wherein x_FFor removing the return value after normalization, x_TmaxAnd x_TminIt is respectively the maximum in test set sample and minima, y_o For network output valve, parameter y_minAnd y_maxValue keep constant, be taken as 0 and 1 respectively.

8. Forecasting Methodology as claimed in claim 1, it is characterised in that in step 10, error analysis expression formula is:

$MAPE=\frac{1}{N}\left(\underset{i=1}{\overset{N}{\Σ}}\right|\frac{{\mathrm{Re}}_i-F_i}{{\mathrm{Re}}_i}|*100\%)---\left(6\right)$

$RMSE=\sqrt{\frac{\underset{i=1}{\overset{N}{\Σ}}{(F_i-{\mathrm{Re}}_i)}^2}{N}}---\left(7\right)$

Wherein Re_iFor actual value, F_iFor predictive value, N is predictive value number；MAPE is mean percent absolute error (Mean Absolute Percentage Error)；RMSE is root-mean-square error (Root Mean Square Error).