CN106022521A - Hadoop framework-based short-term load prediction method for distributed BP neural network - Google Patents

Abstract

The invention discloses a Hadoop framework-based short-term load prediction method for a distributed BP (Back Propagation) neural network. The method specifically comprises the steps of obtaining an initial load data set; dividing the load data set into small data sets and storing the small data sets in data nodes of a distributed file system; initializing BP neural network parameters and uploading a parameter set into the distributed file system; training the BP neural network according to a current load sample, and obtaining correction values of a weight and a threshold of the BP neural network in the current data set; performing statistics on sum of weight and threshold parameters of all layers and between the layers of the network according to a key value of a key value pair; judging whether the convergence precision or the maximum iterative frequency is reached or not in a current iterative task, and if yes, establishing a distributed BP neural network model, or otherwise, performing correction of the weight and threshold parameters of the network; and inputting prediction day data and obtaining load power data of a prediction day. According to the method, the load prediction speed is increased and the requirements of load prediction precision are met.

Description

The short-term load forecasting method of distributed BP neutral net based on Hadoop framework

Technical field

The present invention relates to power system and be combined short-term load forecasting applied technical field under scene with big data, particularly to A kind of short-term load forecasting method of distributed BP neutral net based on Hadoop framework.

Background technology

Load forecast ensure Power System Planning with reliably, have in terms of economical operation and be of great significance. Along with the constantly progress of modern technologies and going deep into of intelligent grid, the development that load prediction theory and technology is the biggest.For many years Coming, Methods of electric load forecasting and theory continue to bring out, and vector is supported in time series method, fuzzy theory, regression analysis, recurrence The technology such as machine, Bayes and neutral net are that load forecast provides good technical support.

Existing algorithm still suffers from certain limitation.Time series method: higher to historical data accuracy, short term is pre- During survey 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.Return Analytic process: describe the quantitative relation between observed variable quantitatively from statistical average meaning visual angle, but very by load data amount The restriction of scale.Regression support vector machine: the method has good generalization ability, meanwhile, also can because of to penalty coefficient c, The e-value of loss function and the optimizing of the γ-value of kernel function and cause the training time the most tediously long, especially advise at training sample set When mould is bigger, embody the most prominent.

Along with emerging in large numbers of intelligent power mass data, existing prediction algorithm cannot meet wanting of predetermined speed and precision of prediction Ask, therefore, it is necessary to find a kind of new method that can meet magnanimity TV university data analysis.BP neutral net has the strongest non- Linear Mapping ability, self-learning capability and fault-tolerant ability, when being applied to load small data set, have precision of prediction higher, The advantages such as training speed is fast.But owing to this algorithm can carry out taking turns training for each load input and output sequence, with Calculating and obtain network each layer weights, the correction of threshold value, when data volume is the biggest, operand will become very big, and unit serial is instructed Practicing the time will be likely to be breached several hours, the most greatly.Therefore, the short-term load forecasting problem on the basis of mass data is still one Individual problem demanding prompt solution.

Summary of the invention

The technical problem to be solved is to provide a kind of distributed BP neutral net based on Hadoop framework Short-term load forecasting method, described method, based on Hadoop cluster platform, plays the nonlinear mapping energy that BP neutral net is powerful Power, improves load prediction speed, meets the requirement of load prediction precision.

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

The short-term load forecasting method of a kind of distributed BP neutral net based on Hadoop framework, comprises the following steps:

Step one: obtain initial load data set；

Step 2: according to the MapReduce framework of Hadoop framework, load data collection is split as small data set, deposits Storage is in the back end of distributed file system；Described small data set key-value pair<key, value>represents, key value is The initial character of this row is relative to the side-play amount of text first address, and value value will be resolvable to the current each layer of BP neutral net Weights, threshold value；

Step 3: initialize BP neural network parameter, comprise the number of plies of input layer, hidden layer and output layer, input layer with The weights between weights, the threshold value of hidden layer neuron, hidden layer and output layer between hidden layer, the threshold of output layer neuron Value, is uploaded to parameter sets in distributed file system；

Step 4: in the Map stage, reads the parameter in distributed file system, including weights, threshold value, appoints at every height When business starts, reduction BP neutral net；The forward of the input signal carrying out BP neutral net according to the distributed data in subtask passes Pass the back propagation with error signal, obtain the weights of BP neutral net, threshold value correction under current data set, according to key It is worth form as the input parameter in Reduce stage；

Step 5: in the Reduce stage, after all data sets are trained by BP neutral net, according to mark input layer, implies Layer and the key value in the key-value pair<key, value>of output layer neuron corresponding weight value and threshold value, all load data samples of statistics This training terminate after to each neuron weights, the influence amount of threshold value, by result output to distributed file system；

Step 6: judge under current iteration task, if reach convergence precision or reach maximum iteration time；If so, depend on According to the number of plies of input layer, hidden layer and the output layer of BP neutral net, and in distributed file system, weights, threshold parameter are built Vertical distributed BP neural network model, if it is not, carry out the correction of BP neural network weight, threshold parameter；

Step 7: according to distributed BP neural network model, input prediction day data be predicted, obtain predicting the negative of day Lotus power data.

According to such scheme, in described step 4, calculate and obtain the weights of BP neutral net, threshold value under current data set Correction, particularly as follows:

Δw′_ki(τ+1)=(1-ρ) η Δ w_ki(τ+1)+ρΔw_ki(τ),

Δ′α_k(τ+1)=(1-ρ) η Δ α_k(τ+1)+ρΔα_k(τ),

Δ′w_ij(τ+1)=(1-ρ) η Δ w_ij(τ+1)+ρΔw_ij(τ),

Δ′θ_i(τ+1)=(1-ρ) η Δ θ_i(τ+1)+ρΔθ_i(τ),

Wherein, Δ w '_kiFor showing that output layer kth node is to the final modified weight amount between hidden layer i-th node； Δ′α_kFinal threshold value correction for output layer kth node；Δ′w_ijRepresent that hidden layer i-th node is to input layer jth Final modified weight amount between node；Δ′θ_iFinal threshold value correction for hidden layer i-th node；ρ is factor of momentum；τ For iterations.

According to such scheme, also include distributed BP neural network model is improved, i.e. introduce factor of momentum, use Repeatedly calculate the mode averaged distributed BP neural network model is improved.

Compared with prior art, the invention has the beneficial effects as follows: use the Hadoop cluster effectively processing the big data of magnanimity Platform, plays the non-linear mapping capability that BP neutral net is powerful, improves load prediction speed, meet load prediction precision Requirement.

Accompanying drawing explanation

Fig. 1 is typical three layers of BP neural network structure schematic diagram.

Fig. 2 is distributed BP neural network prediction model structural representation in the present invention.

Fig. 3 is laboratory Hadoop cluster platform topological diagram in the present invention.

Fig. 4 is that in the present invention, MapReduce-BP load prediction results and traditional BP predict the outcome comparison diagram.

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.Feature of the present invention has three: One, the forward transmission of the input signal of traditional BP neutral net, the back-propagation process of error signal are fully resolved；Two, by BP The training process of neutral net combines with MapReduce framework, studies and passes through Java language realization based on MapReduce frame The distributed BP neural network model of frame, rear abbreviation MapReduce-BP model；Three, introduce factor of momentum, use and repeatedly count The mode averaged, improves BP neutral net and is easily absorbed in the problem of local convergence, improves its resistant and oscillation resistant ability.Describe in detail such as Under:

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 a kind of energy Learn and store substantial amounts of input-output mode map relation, it is not necessary to the math equation in advance disclosing this mapping relations is many Layer feedforward neural network, is made up of input layer, hidden layer and output layer.Fig. 1 is the knot of typical three layers of BP neutral net Composition, 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. In Fig. 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_kRepresent 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 reality Carrying out by the direction being input to output during the output of border, each layer weights, the makeover process of threshold value then enter from exporting to the direction inputted OK.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 forward transmittance 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 o_i, input quantity net of output layer kth node_kWith output o_kIt is respectively

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

$o_i=\mathrm{\&phi;}\left({\mathrm{net}}_i\right)=\mathrm{\&phi;}(\underset{i=1}{\overset{M}{\&Sigma;}}{w}_{ij}{x}_j+{\mathrm{\&theta;}}_j)---\left(2\right)$

${\mathrm{net}}_k=\underset{i=1}{\overset{q}{\&Sigma;}}{w}_{ki}\mathrm{\&phi;}(\underset{i=1}{\overset{M}{\&Sigma;}}{w}_{ij}{x}_j+{\mathrm{\&theta;}}_j)+{\mathrm{\&alpha;}}_k---\left(3\right)$

$o_k=\mathrm{\&psi;}\left({\mathrm{net}}_k\right)=\mathrm{\&psi;}\&lsqb;\underset{i=1}{\overset{q}{\&Sigma;}}{w}_{ki}\&CenterDot;\mathrm{\&phi;}(\underset{i=1}{\overset{M}{\&Sigma;}}{w}_{ij}{x}_j+{\mathrm{\&theta;}}_j)+{\mathrm{\&alpha;}}_k\&rsqb;---\left(4\right)$

(2) back-propagation process of error signal

First by output layer, successively calculate the output error of each layer neuron, then adjust according to error gradient descent method Saving weights and the threshold value of each layer, after making regulation, final output of network mapping can be close to expected value.According to error gradient descent method, Revise hidden layer successively to output layer modified weight amount Δ w_ki, output layer threshold value correction amount α_k, input layer is to hidden layer weights Correction amount w_ijWith hidden layer threshold value correction amount θ_i, as shown in formula (5)-formula (8), in formula, η is learning rate；P is training sample This sum.

${\mathrm{\&Delta;w}}_{ki}=\mathrm{\&eta;}\underset{p=1}{\overset{P}{\&Sigma;}}\underset{k=1}{\overset{L}{\&Sigma;}}(T_k^p-o_k^p){\mathrm{\&psi;}}^{\&prime;}\left({\mathrm{net}}_k\right)o_i---\left(5\right)$

${\mathrm{\&Delta;\&alpha;}}_k=\mathrm{\&eta;}\underset{p=1}{\overset{P}{\&Sigma;}}\underset{k=1}{\overset{L}{\&Sigma;}}(T_k^p-o_k^p){\mathrm{\&psi;}}^{\&prime;}\left({\mathrm{net}}_k\right)---\left(6\right)$

${\mathrm{\&Delta;w}}_{ij}=\mathrm{\&eta;}\underset{p=1}{\overset{P}{\&Sigma;}}\underset{k=1}{\overset{L}{\&Sigma;}}(T_k^p-o_k^p){\mathrm{\&psi;}}^{\&prime;}\left({\mathrm{net}}_k\right)w_{ki}{\mathrm{\&phi;}}^{\&prime;}\left({\mathrm{net}}_i\right)x_j---\left(7\right)$

${\mathrm{\&Delta;\&theta;}}_i=\mathrm{\&eta;}\underset{p=1}{\overset{P}{\&Sigma;}}\underset{k=1}{\overset{L}{\&Sigma;}}(T_k^p-o_k^p){\mathrm{\&psi;}}^{\&prime;}\left({\mathrm{net}}_k\right)w_{ki}{\mathrm{\&phi;}}^{\&prime;}\left({\mathrm{net}}_i\right)---\left(8\right)$

2, distributed BP neural fusion based on MapReduce framework

BP neural network weight, updating of threshold value are not directly dependent upon with the input sequence of sample, therefore, and BP neutral net Training process can use and is divided in by training data on the different PCs of cluster, and each PC is trained for different data Mode, this mode is much like with the principle of MapReduce programming framework.To this end, combine MapReduce framework and BP nerve net Network data can be parallel feature, set up the Distributed Predictive model of BP neutral net based on MapReduce framework, mainly comprise Map stage, Reduce stage and driving function three part.

1) the Map stage

The storage being uploaded to distributed file system HDFS is characterized each layer weights of BP neutral net, the textual data of threshold value Being divided into several data subsets according to set, data subset key-value pair<key, value>represents.Key value is the lead-in of this row According with the side-play amount relative to text first address, value value will be resolvable to current BP neutral net each layer weights, threshold value. This stage carries out forward transmission and the back propagation of error signal of input signal for current loads sample, with obtain weights, The correction of threshold value.Weights herein, the correction of threshold value correspond to the value of the output key-value pair in Map stage, and key value It is each layer weights, the another name of threshold value.Generate weights, threshold value correction Map stage key takeaway as follows:

Input: current loads data sample and each layer weights of BP neutral net, threshold value.

Output: the current weights of sample, the correction of threshold value.

Method: according to the transmission of input signal forward and error signal back propagation, calculate and obtain weights, the correction of threshold value Amount, specific as follows:

(1) setup function

According in HDFS file system, store weights, the text of threshold value, parse current BP by storage principle and order Neutral net each layer weights (w_kiAnd w_ij), threshold quantity (α_kAnd θ_i)。

(2) correction of weights, threshold value

After introducing factor of momentum, each layer weights, the more new formula of threshold value translate into formula (9)-formula (12).Wherein, Δ w '_ki For showing that output layer kth node is to the final modified weight amount between hidden layer i-th node；Δ′α_kSave for output layer kth The final threshold value correction of point；Δ′w_ijFor the final modified weight between hidden layer i-th node to input layer jth node Amount；Δ′θ_iFinal threshold value correction for hidden layer i-th node；ρ is factor of momentum；τ is iterations.Adjust after renewal Formula is specific as follows:

Δw′_ki(τ+1)=(1-ρ) η Δ w_ki(τ+1)+ρΔw_ki(τ) (9)

Δ′α_k(τ+1)=(1-ρ) η Δ α_k(τ+1)+ρΔα_k(τ) (10)

Δ′w_ij(τ+1)=(1-ρ) η Δ w_ij(τ+1)+ρΔw_ij(τ) (11)

Δ′θ_i(τ+1)=(1-ρ) η Δ θ_i(τ+1)+ρΔθ_i(τ) (12)

Formula (1)-formula (4) can be passed through, formula (9)-formula (12) be respectively completed load input signal forward transmission and error believe Number back propagation, obtain under current loads sequence with this, the weights of each layer of BP neutral net, the correction of threshold value.

(3) map function

Map function is mainly in combination with (1) and (2) part, after the calculating completing weights, threshold value correction, uses context Weights, threshold value correction are exported by mode with key-value pair<key, value>, and concrete form is<IntWritable, Text>.

2) the Reduce stage

According to the key value in key-value pair, complete BP neutral net entirety weights, threshold value renewal statistics of variables in conjunction with value. Note 2 points: one, now value form is Iterable<Text>set, set dimension is the total of load sequence training sample Number, therefore needs progressively traversal to obtain correction character string Text of current sample before correction；Two, this stage has been nerve net Network weights, the summation of threshold value correction, summation is the cumulative of numerical value, therefore needs would indicate that the character of the correction of weights, threshold value Illustration and text juxtaposed setting originally resolves to numeric form.Summation degneracy, threshold value correction Reduce stage key takeaway as follows:

Input: what the Map stage exported characterizes single load sequence samples to repairing after BP neural network weight, adjusting thresholds Positive quantity key-value pair<IntWritable, Text>.

Output: BP neutral net entirety weights, the renewal amount of threshold value.

Method: according to BP neutral net each layer weights, the another name of threshold value, is also the key of the input key-value pair in this stage, point Lei Jia not add up overall weights, the renewal amount of threshold value.And the vital point in this stage is in reduce function, successively complete defeated Enter the value of key-value pair, i.e. Iterable<Text>traversal, obtain the weights of current sample, threshold value correction character string and After Text text is converted into numerical morphological, according to key value complete corresponding weight value, threshold value correction cumulative.

3) driving function

Driving function can be regarded as the configuration file of whole program, mainly completes the relevant setting of Job operation.The present invention exists In driving function, mainly complete three relevant settings: one, generate the weights of initial BP neutral net, threshold value text, and by it Reach in HDFS distributed file system；Two, create MapReduce operation, and the map function of Mapper class, run-out key are set The value data type to<key, value>, and the reduce function of Reducer class, the data of output key-value pair<key, value> Type；Three, in view of BP neutral net being the process of an iterative, therefore corresponding homework type is iteration MapReduce Calculating task, therefore needs to arrange iteration ending standard, i.e. sets maximum iteration time and the error margin of training.

Below by instantiation, technical solution of the present invention and technique effect are further described.

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, during the sampling of each equipment Between gap periods be 1h, Weather information is dry-bulb temperature, dew point temperature.Although laboratory data amount is not reaching to the rule of big data Mould, but the correctness of the inventive method can be verified by this experimental data, thus provide for the load prediction under big data environment A kind of new method.Training area is the electricity consumption data on March 31st, 1 day 1 January in 2014.Prediction day is 2014 4 Month electric load the most in the same time on the 1st, 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 with Moment load, the last week day of 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 are together Moment dry-bulb temperature, 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 prediction were revealed the same day in the same time Point temperature, 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 minus Before lotus sequence inputting distributed BP neural network model, needing first to be normalized data, training carries out anti-again after terminating Normalized, obtains actual load predictive value.

2, experimental result and analysis

This experiment is MapReduce-BP neutral net to be compared with traditional BP neural network algorithm, to confirm two Point, one, BP neural network weight, the makeover process of threshold value the most do not associate with the training sequencing of load list entries, I.e. BP neural network training process can be converted into data parallel；Two, demonstrate that to write BP based on MapReduce framework neural The correctness of the thinking of network load forecast model.Fig. 4 is MapReduce-BP load prediction results and traditional BP in the present invention Predict the outcome comparison diagram.Understand, based on the same prediction of prediction curve to prediction daily load of the parallel MapReduce-BP neutral net Day realized load curve is more identical, and traditional BP neural network prediction value prediction effect is close.Wherein, MapReduce-BP bears The average relative error of lotus prediction, root-mean-square error are respectively 3.95% and 1.97%；Traditional BP neural network prediction average Relative error, root-mean-square error are 3.92% and 1.93%.Thus it was confirmed carried MapReduce-BP power load distributing formula is predicted The correctness of model.

Claims (3)

1. the short-term load forecasting method of a distributed BP neutral net based on Hadoop framework, it is characterised in that include Following steps:

Step one: obtain initial load data set；

Step 2: according to the MapReduce framework of Hadoop framework, load data collection is split as small data set, is stored in In the back end of distributed file system；Described small data set key-value pair<key, value>represents, key value is this row Initial character relative to the side-play amount of text first address, value value will be resolvable to current BP neutral net each layer weights, Threshold value；

Step 3: initialize BP neural network parameter, comprises the number of plies of input layer, hidden layer and output layer, and input layer is with implicit The weights between weights, the threshold value of hidden layer neuron, hidden layer and output layer between Ceng, the threshold value of output layer neuron, Parameter sets is uploaded in distributed file system；

Step 4: in the Map stage, reads the parameter in distributed file system, including weights, threshold value, opens in each subtask During the beginning, reduction BP neutral net；Carry out according to subtask distributed data BP neutral net input signal forward transmission with The back propagation of error signal, obtains the weights of BP neutral net, threshold value correction under current data set, and according to key assignments To form as the input parameter in Reduce stage；

Step 5: in the Reduce stage, BP neutral net to all data sets train after, according to mark input layer, hidden layer with Key value in the key-value pair<key, value>of output layer neuron corresponding weight value and threshold value, statistics all load data samples instruction Practice to each neuron weights, the influence amount of threshold value after terminating, by result output to distributed file system；

Step 6: judge under current iteration task, if reach convergence precision or reach maximum iteration time；If so, according to BP The number of plies of the input layer of neutral net, hidden layer and output layer, and weights, threshold parameter in distributed file system, set up Distributed BP neural network model, if it is not, carry out the correction of BP neural network weight, threshold parameter；

Step 7: according to distributed BP neural network model, input prediction day data be predicted, obtain predicting the load merit of day Rate data.

2. the short-term load forecasting method of distributed BP neutral net based on Hadoop framework as claimed in claim 1, its It is characterised by, in described step 4, calculates and obtain the weights of BP neutral net, threshold value correction under current data set, tool Body is:

Δw′_ki(τ+1)=(1-ρ) η Δ w_ki(τ+1)+ρΔw_ki(τ),

Δ′α_k(τ+1)=(1-ρ) η Δ α_k(τ+1)+ρΔα_k(τ),

Δ′w_ij(τ+1)=(1-ρ) η Δ w_ij(τ+1)+ρΔw_ij(τ),

Δ′θ_i(τ+1)=(1-ρ) η Δ θ_i(τ+1)+ρΔθ_i(τ),

Wherein, Δ w '_kiFor the final modified weight amount between output layer kth node to hidden layer i-th node；Δ′α_kFor The final threshold value correction of output layer kth node；Δ′w_ijBetween hidden layer i-th node to input layer jth node Final modified weight amount；Δ′θ_iFinal threshold value correction for hidden layer i-th node；ρ is factor of momentum；τ is iteration time Number.

3. the short-term load forecasting method of distributed BP neutral net based on Hadoop framework as claimed in claim 1 or 2, It is characterized in that, also include distributed BP neural network model is improved, i.e. introduce factor of momentum, use repeatedly to calculate and ask Distributed BP neural network model is improved by the mode of meansigma methods.