CN110070145A - LSTM wheel hub single-item energy consumption prediction based on increment cluster - Google Patents

LSTM wheel hub single-item energy consumption prediction based on increment cluster Download PDF

Info

Publication number
CN110070145A
CN110070145A CN201910364392.2A CN201910364392A CN110070145A CN 110070145 A CN110070145 A CN 110070145A CN 201910364392 A CN201910364392 A CN 201910364392A CN 110070145 A CN110070145 A CN 110070145A
Authority
CN
China
Prior art keywords
energy consumption
lstm
cluster
data
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910364392.2A
Other languages
Chinese (zh)
Other versions
CN110070145B (en
Inventor
陈珊珊
马东方
路海伦
焦正杉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianjin Development Zone Jingnuo Ocean Data Technology Co Ltd
Original Assignee
Tianjin Development Zone Jingnuo Ocean Data Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianjin Development Zone Jingnuo Ocean Data Technology Co Ltd filed Critical Tianjin Development Zone Jingnuo Ocean Data Technology Co Ltd
Priority to CN201910364392.2A priority Critical patent/CN110070145B/en
Publication of CN110070145A publication Critical patent/CN110070145A/en
Application granted granted Critical
Publication of CN110070145B publication Critical patent/CN110070145B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Business, Economics & Management (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Human Resources & Organizations (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Software Systems (AREA)
  • Economics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Strategic Management (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses the LSTM wheel hub single-item energy consumption clustered based on increment predictions, are related to wheel hub single-item energy consumption electric powder prediction.The present invention realizes the clustering to hubless feature parameter using the dynamic increment Density Clustering based on PCA, obtain the affiliated historical product classification of newly-increased product, and single-item energy consumption strong solution is filtered out using Pearson coefficient and Adaptive-Lasso algorithm based on energy consumption Factor system and releases sexual factor, the prediction of new product explanatory variable by force is realized using BP, constructing LSTM energy consumption prediction model for the product that respectively clusters realizes being effectively predicted for new product unit consumption, the optimization to LSTM is realized using ADE, while the dynamic for introducing incremental learning strategy implementation model updates.The present invention demonstrates the validity of prediction technique, make the root-mean-square error RMSE of energy consumption prediction down to 0.016524,0.013653 is averagely reduced relative to other algorithms without incremental learning, the search performance of its ADE is more excellent simultaneously, the RMSE ratio of training set has the average reduction 0.004089 of incremental learning DE-LSTM, and effectively shortens runing time.

Description

LSTM wheel hub single-item energy consumption prediction based on increment cluster
Technical field
The present invention relates to wheel hub single-item energy consumption electric powder prediction more particularly to a kind of LSTM wheel hubs based on increment cluster Single-item energy consumption prediction technique.
Background technique
State's inner wheel hub industry is driven to realize under the thriving demand in vehicle market fast-developing, at present China existing more than 300 Wheel hub manufacturer, family, automotive hub yield keep double-digit growth speed, it is contemplated that domestic automobile after market wheel hub in 2022 Demand is at 6,040,000 or so, it is seen that wheel products have entered " made in China epoch ".Therefore for response market demand and Meet the personalized customization requirement of customer, the production model that wheel hub manufacturing enterprise gradually develops as more category small lots.But working as needs When producing the new product of no knowhow, due to the accumulation of no historical energy consumption data, the unpredictable new product energy consumption of conventional model, And energy consumption directly affects the single-item cost of wheel hub, therefore the energy consumption for studying wheel hub new product is of great significance.
At present in terms of establishing energy consumption prediction model, because of the depth integration of intelligent production and technology of Internet of things, industry is raw Data acquisition facilitates quick during production, and various intelligent predicting technologies have been used for energy demand management, not with Accurate Prediction The energy demand come is mainly taken using time series forecasting, gray prediction, support vector machines, neural network prediction etc. at present Obtained preferable research achievement.Article [Air Transport Enterprises Under energy consumption prediction model (English) of the such as Geng Hong based on small echo-ARIMA [J] lathe and hydraulic, 2018,46 (06): 13-17+42.] propose that a kind of ARIMA decomposed based on two multi-scale wavelets predicts mould Type；Article [Su Wei equalization papermaking enterprise technical process energy consumption predictive simulation [J] Computer Simulation, 2016,33 (08): 438-442+447.] by being improved to the maximum likelihood parameter estimation method in ARIMA model, it proposes based on Bayes's ARIMA model carries out energy consumption prediction.[such as Li Lixin are based in energy consumption prediction [J] of Gray Markov Model for article State's scientific and technological information, 2018 (15): 74-75.] it is directed to the prediction of AND ENERGY RESOURCES CONSUMPTION IN CHINA total amount, propose the base in grey forecasting model The viewpoint of Markov chain is introduced on plinth；Article [boat of the such as Liu Jiaxue based on metabolic grey Markov-arma modeling Empty company's energy consumption prediction (English) [J] lathe and hydraulic, 2017,45 (18): 55-62.] by the prediction result of GM (1,1) by horse Er Kefu amendment, and with metabolism method deleting madel in no longer effective property legacy data, by time slip-window with Arma modeling is to residual error corrections；Article [application of the .GM-WLSSVM such as Du Ruizhi model in the prediction of office building electric power energy consumption [J] computer application and software, 2018,35 (09): 44-49+55.] utilize gray model to choose different samples to same a period of time Prediction result, is combined followed by Weighted Least Squares Support Vector Machines model, realizes by the prediction of Duan Jinhang polymorphic type The short-term forecast of office building electric energy；[such as Wang Kun are predicted article based on the airport energy consumption of the EMD and LSSVM of drosophila parameter optimization [J] computer age, 2017 (04): 35-40.] propose the minimum two of a kind of combination empirical mode decomposition and drosophila parameter optimization Multiply the energy consumption prediction technique of support vector machines；Article [papermaking enterprise energy consumption prediction model of the such as summer prestige an ancient unit of weight based on PSO-LSSVR Research [J] computer measurement and control, 2013,21 (12): 3433-3435+3438.] propose it is a kind of based on particle group optimizing Least square support vector regression energy consumption prediction model；[it is pre- that Guo Xiao waits the short-term energy optimization of the Civil Aviation Airport electricity consumption quietly to article Survey emulation [J] Computer Simulation, 2018,35 (09): 31-36.] propose improved grey color depth conviction net combined prediction mould Type improves the short-term energy consumption model precision of prediction of Civil Aviation Airport electricity consumption；[such as Chen Zhoulin improve PSO-BP neural network forecast mould to article Application [J/OL] the light science and technology of type in the prediction of papermaking energy consumption, 2018 (11): 91-94 [2019-01-08.], which is established, to be based on Improve the BP neural network energy consumption prediction model of particle swarm algorithm optimization；[such as Chen Yanxi are built article based on the green offfice of ANN Build HVAC system operation energy consumption prediction [J] building energy conservation, 2017,45 (10): 1-5.] establish classification multilayer perceptron nerve Network Prediction Model realizes the energy consumption prediction of heating ventilation air-conditioning system in office building；[the one kind such as Zhang Yuhang is based on article Short-term electro-load forecast method [J] the power information and the communication technology of LSTM neural network, 2017,15 (09): 19-25.] Using Power system load data itself as training data and output label, the electricity based on LSTM is established by the method for repetitive exercise Power load forecasting model.
It can be seen that from above-mentioned document currently without the forecasting problem for researching and analysing single-item energy consumption, and hybrid prediction model More it is capable of the overall performance of improved model.Due to the relationship between energy consumption and its influence factor be it is nonlinear, using nerve net Network more can be realized accurate prediction, but the time modeling ability of traditional feedforward neural network is fairly limited, in output In the case that predicted value depends on the long history of input feature vector sequence, LSTM is able to solve the problem of learning long-rang dependence.And such as The strong explanatory influence factor what obtains energy consumption is the difficult point of research and the basis of building prediction model, and is pushed away along with the time It moves, data are increasing, these above-mentioned prediction techniques are conventional batch predictions, be not suitable for the constantly newly-increased scene of data volume, How incremental learning and current urgent problem carried out on original learning foundation.
Summary of the invention
The purpose of the present invention is to provide a kind of LSTM wheel hub single-item energy consumption prediction techniques based on increment cluster.It is intended to adopt With PCA to the feature vectors dimensional down for determining history wheel hub production model after, obtained and similar with new product gone through using clustering algorithm History single-item classification；Then, energy consumption Factor system is constructed based on order data, creation data, crucial consumable component parameter, utilized Pearson coefficient and Adaptive-Lasso obtain with the strongly connected strong explanatory variable of single-item energy consumption, and use BP nerve Sexual factor value is released in the strong solution that neural network forecast goes out new product；The single-item energy for the wheel hub that respectively clusters finally is constructed based on the above method Prediction model is consumed, proposes that a kind of LSTM incremental update wheel hub unit consumption prediction model based on ADE, the model use ADE algorithm pair The initial parameter of LSTM model scans for, and when there is newly-increased sample, is updated to prediction model.The method achieve right The incrementally updating of new wheel hub single-item energy consumption being effectively predicted with model, not only increases precision of prediction, while reducing data Memory space and model calculate the time.
The technical solution used in the present invention is a kind of LSTM wheel hub single-item energy consumption prediction technique based on increment cluster, It is characterized in that including the following steps:
(1) collect the characteristic parameter sample of different model wheel hub, including 12 characteristic parameters: wheel diameter, wheel rim width, Centre bore away from, bolt hole count, pitch diameter, offset distance, weight, spoke number, spoke front pose, center disk moulding, wheel hub Material, manufacturing process；
(2) order data, creation data, the crucial consumable component parameter, single-item energy consumption data composition of different model wheel hub are collected Energy consumption sample；Wherein order data, creation data, crucial consumable component parameter constitute energy consumption Factor system, order data packet Single-item model, single-item order volume, category sum are included, when creation data includes production time, raw material input amount, equipment operation Between, production efficiency, rejection rate, wherein production efficiency by machine add efficiency (part/hour) and the aspect of finishing efficiency (minute/part) two Lai Consider, crucial consumable component parameter includes drill bit usage amount, cutter usage amount；
(3) after the character shape parameter of history hubless feature parameter sample in step (1) being carried out labeling processing, using master Parameter sample is down to two dimension and obtains data set P by constituent analysis；
(4) cluster operation then is carried out to P using dynamic increment density clustering algorithm, obtains the original result C that clustersi(i= 0,1 ..., k) it ∈ P and peels off and collects O ∈ P；
(5) when there is newly-increased feature parameter sample, to its using step (3) pretreatment after Δ P, search for O ∪ Δ P in Original clusters the reachable data object of density, updates cluster result, and output clusters C 'i(i=0,1 ..., k ')=Ci∪ΔCi, peel off Collect O ' and newly-increased product generic；
(6) sexual factor is released according to the strong solution of energy consumption Factor system analysis single-item energy consumption in step (2), utilized Pearson coefficient rejects the weak related and uncorrelated factor in energy consumption Factor system, followed by Adaptive- Lasso algorithm carries out second of variables choice and obtains the strong explanatory variable of wheel hub single-item energy consumption；
(7) strong explanatory to being obtained by step (3) labeling treated history hubless feature parameter, step (6) Variable and single-item power consumption values are standardized；
It (8) is input with the history hubless feature parameter after step (7) standardization, to be marked by step (7) Treated that strong explanatory variable is output for standardization, constructs the BP prediction model of new wheel hub explanatory variable by force, it is new to be able to prediction The strong explanatory variate-value of wheel hub；
(9) the single-item energy consumption prediction model that respectively clusters is constructed according to the Historic Clustering result of step (4), by the history energy that respectively clusters It consumes sample standard deviation and training set and test set is randomly divided into the ratio of 3:1, training set is divided into four groups, wherein one group of carry out step (9), remaining set is successively used as a sample increment collection；
(10) the LSTM single-item energy consumption prediction model that respectively clusters is constructed, by training set after step (7) standardization Strong explanatory variable is input, and the single-item energy consumption after step (7) standardization is output, is calculated in model construction using ADE Method realizes the optimization to LSTM parameter；
(11) it in the LSTM single-item energy consumption prediction model that respectively clusters, is carried out using sample increment the set pair analysis model in step (9) Incremental update；
(12) according to the test set test respectively to cluster in step (9) by step (11) updated prediction mould respectively to cluster Type, and model evaluation is carried out, it is handled simultaneously for the characteristic parameter sample of product to be predicted by step (3)-(5), Yi Jijing The strong explanatory variable that step (7)-(8) BP prediction model obtains new wheel hub is crossed, it is updated each followed by step (11) The LSTM single-item energy consumption prediction model that clusters is predicted, the predicted value of newly-increased product energy consumption is exported.
A further technical solution lies in cluster the characteristic parameter sample set P after dimensionality reduction in the step (4) Process includes the following steps:
1) for the raw data set P after dimensionality reduction, determine the value of coefR, calculate density adjusting parameter σ and density up to away from From R, and calculate the density value Density (P of each data object in Pi), obtain Density (P in Pi) maximum local density Attractor Attractor:
In formula,Indicate point PiTo point PjEuclidean distance,The average value of distance between each point in sample, CoefR (0 < coefR < 1) is the original regulation coefficient of density reach distance, and n is total sample number；
2) data object in raw data set P is scanned, by density-attractors Attractor and the reachable data pair of its density As being assigned to first cluster C0, and concentrated from initial data and delete the object that clusters；
3) for remaining data collection, another density-attractors Attractor is searchedi, calculate density self-adapting up to away from From Radap,i, the density-attractors and the reachable data object of its density are assigned to another C that clustersi, and from raw data set The middle deletion object that clusters, and so on, finally the cluster with little data object is put into exceptional value or noise group, at this time Obtain the original result C that clustersi(i=0,1 ..., k) ∈ P, peels off and collects O ∈ P:
Radap,i=α R
α is regulation coefficient in formula, and formula is as follows:
A further technical solution lies in increment clustering methods in the step (5) are as follows:
1) for pretreated incremental data set Δ P, Δ P, which is peeled off with original, to be collected O and merges, and is searched for each with original in O ∪ Δ P Cluster CiThe reachable data object of density updates cluster result Ci(i=0,1 ..., k) ∈ (P ∪ Δ P), at this time remaining data collection For
2) to remaining data collectionData object carry out clustering, be likely to be present inIn cluster Δ Ci(i=k+1 ..., k ', k ' >=k), update cluster result as C 'i(i=0,1 ..., k ') =Ci(i=0,1 ..., k) ∪ Δ Ci, and peel off and collect O '；
If 3)It then detects whether containing the data object in increment sample Δ P in O ', if joining containing new product Numerical example, then the distance for comparing the new sample point to the middle each point that respectively clusters will in order to realize the energy consumption prediction of the new product It is first tagged to apart from cluster where its closest approach；
4) after for the pretreatment of other incremental data sets, the existing label that clusters collected in O ' that peels off is deleted, step is repeated It is rapid 1) to 3)；
5) cluster result C ' is exportedi(i=0,1 ..., k ') and peel off set O ' and new product generic.
A further technical solution lies in single-item energy consumption prediction model construction methods in the step (10) are as follows:
1) dimension of ADE individual is equal to the weight of LSTM neural network and the summation of threshold number, initializes evolution number G =0, population scale N, crossover probability f, mutation probability cr, and using root-mean-square error RMSE as fitness function:
GenM is maximum number of iterations in formula, and G is current iteration number, and α, β are the constant in [0.5,1] range, ytFor True value, k are the number of data object；
2) fitness value, that is, RMSE of each individual in population is calculated；
3) if minimum RMSE is met the requirements in current population or current iteration number G=GenM, ADE termination changes In generation, obtains optimized individual and executes step 5), otherwise continues to execute step 4)；
4) new group is obtained according to adaptive intersection, TSP question and selection operation, G=G+1 is set, return executes step It is rapid 3)；
5) optimize the initial connection weight and threshold value for obtaining optimum individual as LSTM based on ADE；And with training sample pair LSTM implements training, and then obtains optimum network.
A further technical solution lies in increase sample newly to the update method of LSTM prediction model in the step (11) Are as follows:
1) for each newly-increased energy consumption sample, first which cluster c judgement belongs toi, corresponding i-th of the LSTM model of i-th of cluster；It determines After affiliated model, model parameter is updated on the basis of model history data training, i.e., by the parameter after historical data training Network is initialized, the strong explanatory variable of new samples is input in prediction model, is obtained newly by the forward calculation of LSTM The predicted value of sampleBy predicted valueIt is added to the error of actual value y on original error function J (θ)；
In formula, parameter θ=(Wf,Wi,Wc,Wo,bf,bi,bc,bo)
2) model parameter value is updated according to error function J (θ) backpropagation of update:
θ '=(Wf-λ*Δwf,…,Wo-λ*Δwo,bf-λ*Δbf,…,bo-λ*Δbo)
In formula, λ is learning rate, Δ wf、ΔwoWith Δ bf、ΔboIt is the weight of neuron and the gradient square of offset respectively Battle array and vector.
The beneficial effects of adopting the technical scheme are that
LSTM wheel hub single-item energy consumption prediction technique proposed by the present invention based on increment cluster, using the dynamic based on PCA Increment Density Clustering realizes the clustering to hubless feature parameter, thus obtain the affiliated historical product classification of newly-increased product, And single-item energy consumption strong solution is filtered out using Pearson coefficient and Adaptive-Lasso algorithm based on energy consumption Factor system and is released Sexual factor, and then realize using BP the prediction of new product explanatory variable by force, it is finally the product building LSTM energy consumption prediction that respectively clusters Being effectively predicted for new product unit consumption of model realization, using ADE realizes the optimization to LSTM, while it is real to introduce incremental learning strategy The dynamic of existing model updates.By analysis of experiments, the LSTM wheel hub single-item energy consumption prediction based on increment cluster of proposition is demonstrated The validity of method, the root-mean-square error RMSE for predicting energy consumption is down to 0.016524, relative to other calculations without incremental learning Method averagely reduces 0.013653, while the search performance of its ADE is more excellent, and the RMSE ratio of training set has incremental learning DE-LSTM Average reduction 0.004089, and effectively shorten runing time.
Detailed description of the invention
The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
Fig. 1 is that the present invention is based on the LSTM wheel hub single-item energy consumption prediction technique structure charts of increment cluster；
Fig. 2 is the dynamic increment Density Clustering flow chart the present invention is based on PCA；
Fig. 3 is that the present invention is based on the LSTM incremental update wheel hub unit consumption of ADE to predict flow chart；
Fig. 4 is cluster analysis result figure of the present invention；
Fig. 5, Fig. 6 are respectively the RMSE variation of the RMSE variation diagram and hidden layer number of nodes and X9 of hidden layer number of nodes and X8 Figure；
Fig. 7 is the RMSE variation diagram of the LSTM number of plies with the prediction model that respectively clusters；
Fig. 8 is the RMSE variation diagram of LSTM node layer number with the prediction model that respectively clusters；
Fig. 9 is cluster C0Energy consumption incremental data in the training precision comparison diagram of five kinds of algorithms；
Figure 10 is cluster C1Energy consumption incremental data in the training precision comparison diagram of five kinds of algorithms；
Figure 11 is cluster C4Energy consumption incremental data in the training precision comparison diagram of five kinds of algorithms；
Specific embodiment
The present invention proposes a kind of dynamic increment density clustering algorithm based on PCA, i.e., based on symbol wheel hub production model Characteristic parameter data obtain history single-item classification similar with new product using clustering algorithm；Then, Pearson coefficient is utilized Sexual factor is released in strong solution with Adaptive-Lasso algorithm analysis single-item energy consumption, and predicts new product using BP neural network Factor value is released in strong solution；Finally, proposing a kind of LSTM incremental update wheel hub unit consumption prediction model based on ADE, which utilizes ADE Algorithm reduces influence of the initiation parameter to model accuracy, and introduces incremental learning strategy, and the dynamic of implementation model updates.
One, the theoretical foundation of the method for the present invention
1, principal component analysis (Principal Component Analysis, PCA): by linear transformation by initial data It is transformed to the expression of one group of each dimension linear independence (unit is orthogonal), can be used for extracting the main feature component of data, protect simultaneously Changeability as much as possible is stayed, the dimensionality reduction of high dimensional data is usually used in.
2, the dynamic increment cluster based on density: it is density-attractors that it is maximum, which to define data dot density, never complete every time It is concentrated at the remaining data of cluster and finds local density's attractor and the reachable sample point of its density, to gather for one kind, and be It is suitable for dynamic incremental data library, it is adaptive to adjust density reach distance.
3, Pearson correlation coefficients (Pearson coefficient): for the degree of correlation reflected between two variables, in engineering It can be used to calculate the similarity between feature and classification in habit, that is, can determine whether that extracted feature and classification are to be positively correlated, is negative Correlation is again without degree of correlation.
4, Adaptive-Lasso algorithm: being the improvement to Lasso algorithm, by construct a penalty function obtain one compared with For the model of refining, so that it compresses some coefficients, concurrently setting some coefficients is zero, and then is realized when variable is numerous Significant variable is fast and effeciently extracted, thus simplified model.
5, shot and long term Memory Neural Networks (Long-Short Term Memory, LSTM): it is the deformation to RNN, that is, exists On the basis of common RNN, by increasing memory unit in each neural unit of hidden layer, to make the recall info in time series It controllably, can by several controllable gates (forgeing door, input gate, candidate door, out gate) when being transmitted between hidden layer each unit every time With the memory of information and current information before control and degree is forgotten, so that RNN network be made to have long-term memory function.
6, adaptive differential evolution algorithm: (Adaptive differential evolution algorithm, ADE): By adaptively adjusting zoom factor and crossover probability in calculating process, the improvement to standard difference evolution algorithm is realized.
Two, based on the LSTM wheel hub single-item energy consumption prediction technique design of increment cluster:
LSTM wheel hub single-item energy consumption prediction technique working principle based on increment cluster is as follows: real to newly-increased product when needing When applying the prediction of single-item energy consumption, after being down to two dimension to the characteristic parameter for determining history wheel hub production model using PCA first, using dynamic State increment density clustering algorithm obtains history single-item classification similar with new product；Then, based on order data, creation data, Crucial consumable component parameter constructs energy consumption Factor system, is obtained and list using Pearson coefficient and Adaptive-Lasso algorithm The strongly connected strong explanatory variable of product energy consumption, and using hubless feature parameter as input, strong explanatory variable as output, building The BP prediction model of new product explanatory variable by force；Finally, constructing the single-item energy consumption for the wheel hub that respectively clusters based on the above method Prediction model weakens influence of the initiation parameter to model using ADE algorithm, it is contemplated that the processing problem of newly-increased energy consumption sample, The strategy for introducing incremental learning, by judging the determination prediction model to be updated that clusters belonging to new samples, by determined model Initial parameter of the history parameters as the model, and the error between new samples predicted value and actual value is added to whole mistake In difference, model parameter is updated come iteration using the method that minimizes the error, guarantees that model can constantly handle the newly-increased of addition with this Sample.
1, the LSTM wheel hub single-item energy consumption prediction technique structure based on increment cluster
The present invention show that history similar with newly-increased production mode produces by the dynamic increment Density Clustering based on PCA Category is other, and is LSTM incremental update wheel hub unit consumption prediction model of the building based on ADE that respectively cluster based on the result that clusters, and realizes single The prediction of product energy consumption.As shown in Figure 1, after it is down to two dimension by PCA, passing through when there is newly-increased product feature parameter sample Dynamic increment density clustering algorithm updates cluster result and obtains its affiliated historical product classification, at the same by Pearson coefficient and Adaptive-Lasso algorithm realizes the screening to energy consumption explanatory variable by force, and predicts to obtain new volume increase by BP neural network Sexual factor is released in the strong solution of product, and then is input to the LSTM model of corresponding A/D E optimization, to obtain the single-item energy consumption of newly-increased product Value, and when there is newly-increased energy consumption sample, according to the error update model error function between its predicted value and true value, adjust mould Shape parameter, the incremental update of implementation model.
2, algorithm implements
LSTM wheel hub single-item energy consumption prediction technique based on increment cluster includes the following steps:
(1) the characteristic parameter sample of different model wheel hub, including 12 characteristic parameters: wheel diameter (p1), wheel rim are collected Width (p2), centre bore are away from (p3), bolt hole count (p4), pitch diameter (p5), offset distance (p6), weight (p7), spoke number (p8), spoke front pose (p9), center disk moulding (p10), hub material (p11), manufacturing process (p12)；
(2) order data, creation data, the crucial consumable component parameter, single-item energy consumption data composition of different model wheel hub are collected Energy consumption sample；Wherein order data, creation data, crucial consumable component parameter constitute energy consumption Factor system, order data packet It is total (X3) to include single-item model (X1), single-item order volume (X2), category, creation data includes production time (X4), raw material throwing Enter amount (X5), operation hours (X6), production efficiency (machine adds efficiency X7, finishing efficiency X8), rejection rate (X9), crucial consumable component Parameter includes drill bit usage amount (X10), cutter usage amount (X11)；
(3) after the character shape parameter of history hubless feature parameter sample in step (1) being carried out labeling processing, using master Parameter sample is down to two dimension and obtains data set P by constituent analysis；
(4) cluster operation then is carried out to P using dynamic increment density clustering algorithm, obtains the original result C that clustersi(i= 0,1 ..., k) it ∈ P and peels off and collects O ∈ P；
(5) when there is newly-increased feature parameter sample, to its using step (3) pretreatment after Δ P, search for O ∪ Δ P in Original clusters the reachable data object of density, updates cluster result, and output clusters C 'i(i=0,1 ..., k ')=Ci∪ΔCi, peel off Collect O ' and newly-increased product generic；
(6) sexual factor is released according to the strong solution of energy consumption Factor system analysis single-item energy consumption in step (2), utilized Pearson coefficient rejects the weak related and uncorrelated factor in energy consumption Factor system, followed by Adaptive- Lasso algorithm carries out second of variables choice and obtains the strong explanatory variable of wheel hub single-item energy consumption；
(7) strong explanatory to being obtained by step (3) labeling treated history hubless feature parameter, step (6) Variable and single-item power consumption values are standardized；
It (8) is input with the history hubless feature parameter after step (7) standardization, to be marked by step (7) Treated that strong explanatory variable is output for standardization, constructs the BP prediction model of new wheel hub explanatory variable by force, it is new to be able to prediction The strong explanatory variate-value of wheel hub；
(9) the single-item energy consumption prediction model that respectively clusters is constructed according to the Historic Clustering result of step (4), by the history energy that respectively clusters It consumes sample standard deviation and training set and test set is randomly divided into the ratio of 3:1, training set is divided into four groups, wherein one group of carry out step (9), remaining set is successively used as a sample increment collection；
(10) the LSTM single-item energy consumption prediction model that respectively clusters is constructed, by training set after step (7) standardization Strong explanatory variable is input, and the single-item energy consumption after step (7) standardization is output, is calculated in model construction using ADE Method realizes the optimization to LSTM parameter；
(11) it in the LSTM single-item energy consumption prediction model that respectively clusters, is carried out using sample increment the set pair analysis model in step (9) Incremental update；
(12) according to the test set test respectively to cluster in step (9) by step (11) updated prediction mould respectively to cluster Type, and model evaluation is carried out, it is handled simultaneously for the characteristic parameter sample of product to be predicted by step (3)-(5), Yi Jijing The strong explanatory variable that step (7)-(8) BP prediction model obtains new wheel hub is crossed, it is updated each followed by step (11) The LSTM single-item energy consumption prediction model that clusters is predicted, the predicted value of newly-increased product energy consumption is exported.
In embodiment of the present invention, the characteristic parameter sample set P after dimensionality reduction clustered in the step (4) Journey includes the following steps:
1) for the raw data set P after dimensionality reduction, determine the value of coefR, calculate density adjusting parameter σ and density up to away from From R, and calculate the density value Density (P of each data object in Pi), obtain Density (P in Pi) maximum local density Attractor Attractor:
In formula,Indicate point PiTo point PjEuclidean distance,The average value of distance between each point in sample, CoefR (0 < coefR < 1) is the original regulation coefficient of density reach distance, and n is total sample number；
2) data object in raw data set P is scanned, by density-attractors Attractor and the reachable data pair of its density As being assigned to first cluster C0, and concentrated from initial data and delete the object that clusters；
3) for remaining data collection, another density-attractors Attractor is searchedi, calculate density self-adapting up to away from From Radap,i, the density-attractors and the reachable data object of its density are assigned to another C that clustersi, and from raw data set The middle deletion object that clusters, and so on, finally the cluster with little data object is put into exceptional value or noise group, at this time Obtain the original result C that clustersi(i=0,1 ..., k) ∈ P, peels off and collects O ∈ P:
Radap,i=α R
α is regulation coefficient in formula, and formula is as follows:
In embodiment of the present invention, increment clustering method in the step (5) are as follows:
1) for pretreated incremental data set Δ P, Δ P, which is peeled off with original, to be collected O and merges, and is searched for each with original in O ∪ Δ P Cluster CiThe reachable data object of density updates cluster result Ci(i=0,1 ..., k) ∈ (P ∪ Δ P), at this time remaining data collection For
2) to remaining data collectionData object carry out clustering, be likely to be present inIn cluster Δ Ci(i=k+1 ..., k ', k ' >=k), update cluster result as C 'i(i=0,1 ..., k ') =Ci(i=0,1 ..., k) ∪ Δ Ci, and peel off and collect O '；
If 3)It then detects whether containing the data object in increment sample Δ P in O ', if joining containing new product Numerical example, then the distance for comparing the new sample point to the middle each point that respectively clusters will in order to realize the energy consumption prediction of the new product It is first tagged to apart from cluster where its closest approach；
4) after for the pretreatment of other incremental data sets, the existing label that clusters collected in O ' that peels off is deleted, step is repeated It is rapid 1) to 3)；
5) cluster result C ' is exportedi(i=0,1 ..., k ') and peel off set O ' and new product generic.
In embodiment of the present invention, single-item energy consumption prediction model construction method in the step (9) are as follows:
1) dimension of ADE individual is equal to the weight of LSTM neural network and the summation of threshold number, initializes evolution number G =0, population scale N, crossover probability f, mutation probability cr, and using root-mean-square error RMSE as fitness function:
GenM is maximum number of iterations in formula, and G is current iteration number, and α, β are the constant in [0.5,1] range, ytFor True value, k are the number of data object；
2) fitness value, that is, RMSE of each individual in population is calculated；
3) if minimum RMSE is met the requirements in current population or current iteration number G=GenM, ADE termination changes In generation, obtains optimized individual and executes step 5), otherwise continues to execute step 4)；
4) new group is obtained according to adaptive intersection, TSP question and selection operation, G=G+1 is set, return executes step It is rapid 3)；
5) optimize the initial connection weight and threshold value for obtaining optimum individual as LSTM based on ADE；And with training sample pair LSTM implements training, and then obtains optimum network.
Sample is increased in embodiment of the present invention, in the step (11) newly to the update method of LSTM prediction model are as follows:
1) for each newly-increased energy consumption sample, first which cluster c judgement belongs toi, corresponding i-th of the LSTM model of i-th of cluster；It determines After affiliated model, model parameter is updated on the basis of model history data training, i.e., by the parameter after historical data training Network is initialized, the strong explanatory variable of new samples is input in prediction model, is obtained newly by the forward calculation of LSTM The predicted value of sampleBy predicted valueIt is added to the error of actual value y on original error function J (θ)；
In formula, parameter θ=(Wf,Wi,Wc,Wo,bf,bi,bc,bo)
2) model parameter value is updated according to error function J (θ) backpropagation of update:
θ '=(Wf-λ*Δwf,…,Wo-λ*Δwo,bf-λ*Δbf,…,bo-λ*Δbo)
In formula, λ is learning rate, Δ wf、ΔwoWith Δ bf、ΔboIt is the weight of neuron and the gradient square of offset respectively Battle array and vector.
Three, data describe
The data of experimental data source Mr. Yu's wheel hub Manufacturing Co., Ltd, the said firm have been respectively mounted Intelligent Instrument in production line Device, for obtaining consumed electric energy, water and natural gas in production process.This is tested for predicting single-item power consumption, Since amount of available data is less, partial data is simulated with Python, has obtained in January, 2015 at present into December, 2018 The related data of 725 wheel-type, in this experiment, wherein 475 wheel-type are used as historical product sample for random selection, remaining 250 wheel-type are considered as newly-increased product for verifying the prediction result of model.
By the characteristic parameter sample of 475 selected wheel-type as raw data set P, the ginseng of remaining 250 wheel-type The random equivalent of numerical example is divided into five groups, for being incremental data set Δ P1、ΔP2、ΔP3、ΔP4、ΔP5Increment cluster is carried out, Specific features parameter pattern representation is as shown in table 1 below.
The description of 1 characteristic parameter data of table
On the basis of cluster result, the LSTM incremental update wheel hub unit consumption prediction model based on ADE is constructed.725 wheels Order data, manufacturing parameter, crucial consumable component parameter and the energy consumption data of type form energy consumption sample, and wherein energy consumption data is single-item Power consumption amounts to 34,800 samples, wherein 475 historical products include 22,800 samples.Historical product is respectively clustered institute Containing sample standard deviation according to 3:1 ratio random division be training sample and test sample, construct the energy consumption prediction model of each cluster. In the training process, the training set equivalent respectively to cluster is divided into four groups, wherein one group for constructing prediction model, remaining three Group is added to existing model and carries out incremental learning.Specific energy consumption data and its Correlative Influence Factors are as shown in table 2 below.
2 energy consumption sample of table
1, model parameter and structure
(1) clustering
For feature spoke front pose (p9), center disk moulding (p10), hub material (p11), manufacturing process (p12) it is character type, needs its standard makeup being changed to numeric type data.
It uses the model split of random sampling at historical product sample in sample, increase product sample newly, wherein historical product 475, remaining is that newly-increased product is added to Clustering Model progress incremental update in five times, and coefR value is 0.6, this time Experiment has carried out 20 random samplings, wheel hub finally can be divided into 5 classes, cluster result is as shown in Figure 4.
One such experimental result is selected to be illustrated: by the parameter sample after randomly select 475 wheel-type dimensionality reductions Clustering is carried out, sample has finally been divided into 6 class C-1、C0、C1、C2、C3、C4, there is 13,261,53,27,32,89 samples respectively Point, wherein C-1It is the cluster comprising noise spot, by C-1Collect O as peeling off.Remaining 250 hub parameter samples point 5 times to cluster As a result incremental update is carried out, searching for can be with the C that clusters in increment sample0、C1、C2、C3、C4The reachable data object of density, finally Cluster C0、C1、C4Separately include 303,78,185 samples, peel off collect O include 9 sample points, C2、C3No new samples are added.
(2) energy consumption Analysis of key influential factors
1) correlation analysis
Tentatively judge whether there is linear dependence, meter between single-item power consumption and each influence factor using Pearson coefficient It calculates shown in result table 3.
3 Person coefficient table of table
X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11
-0.015 0.804 0.637 0.003 0.004 0.612 -0.896 0.748 0.530 0.629 0.616
It follows that X1, X4, X5 and single-item energy consumption are almost uncorrelated, X7 and single-item energy consumption are in strong negatively correlated, surplus variable It is rendered as being positively correlated.Therefore by being somebody's turn to do as a result, these three weak correlative factors of X1, X4, X5 can be removed.
2) Adaptive-Lasso variables choice
Then second of Variable Selection is carried out using Adaptive-Lasso algorithm to surplus variable, as a result shown in table 4:
4 Adaptive-Lasso coefficient table of table
X2 X3 X6 X7 X8 X9 X10 X11
-0.1705 0.0776 0.0000 0.0000 0.4209 0.3929 0.0000 0.0000
It is found that the coefficient of X6, X7, X11, X10 are 0, then when constructing model, these variables are removed, surplus variable X2, X3, X8, X9 are then used as the key influence factor of single-item power consumption.
3) sexual factor prediction is released in the strong solution of new product
By above-mentioned analysis, show that X2, X3, X7, X9 are that sexual factor is released in the strong solution of energy consumption, X2 is order volume, and X3 is same The different product model number that production line produces within the same production cycle, therefore can be according to reality for X2, X3 for increasing product newly Situation is manually set, and X8, X9 need to obtain using BP neural network prediction, becomes 12 characteristic parameters of wheel hub as input Amount, X8, X9 are to export.
The mode that experimentation is all made of k-fold cross validation is trained model, and initial data is divided into k group (K- Fold), each subset data is made into one-time authentication collection respectively, remaining k-1 group subset data obtains k mould as training set Type.This k model is respectively in verifying central evaluation as a result, the smallest model of final choice error.K is taken as 4 in an experiment, i.e., will Data are divided into 4 parts, and k-1=3 parts of data therein is selected to train, and remaining a for verifying, network is 4 times trained in total, often A data have the opportunity to as checksum set.Cycle-index trained every time is gradually increased to 300 during the experiment, i.e., Epoch=300, and getting off after each training to the error log of verification data.In each training process, by learning rate Dynamic adjustment is carried out, when learning to stagnate, learning rate is less in 0.1 times of form.
It is identical by setting BP other parameters for the determination of hiding number of layers, hiding number of layers is gradually increased by 1 layer To 5 layers test predicted value RMSE, as shown in table 5 below, hidden layer be 1 layer when X8, X9 RMSE value it is smaller, when hide layer by layer When number is gradually increased to 5 by 2, the value of RMSE fluctuates in smaller range, therefore selects 1 hidden layer.
Table 5 hides influence of the number of plies to key influence factor precision of prediction
For the determination of hidden layer interstitial content, same setting BP other parameters are identical, and according to formulaSearch, x be input layer number i.e. 12, y be output layer number of nodes i.e. 2, a be it is normal more than or equal to 1 Number, z is hidden layer number of nodes, therefore can set 5 for hidden layer number of nodes initial value and gradually increase, and records the mistake of predicted value Difference, experimental result is as shown in Figure 5,6, and when hiding node layer increases to 16 by 5, the RMSE of finishing efficiency and rejection rate is gradually It reduces, and when hidden layer number of nodes continues growing, the RMSE of finishing efficiency and rejection rate is gradually increased, therefore hides node layer Number is taken as 16.
By above-mentioned experimental result it is found that predicting that the BP neural network of newly-increased product key influence factor is total to 3-tier architecture, wherein Input layer number identical as hubless feature parameter dimension i.e. 12, hidden layer number of nodes is 16, and output layer number of nodes is 2.By this Model running 10 times, the average value of the predicted value of obtained X8, X9 does final predicted value, while each newly-increased product is arranged X2, X3 variable, for doing the input of verifying energy consumption prediction model.
(3) respectively cluster energy consumption prediction model
According to above-mentioned clustering as a result, historical product has been aggregated into 5 classes, historical product is respectively clustered contained energy consumption Sample standard deviation is training sample and test sample according to the ratio random division of 3:1, constructs the energy consumption prediction model of each cluster.It is instructing During white silk, the training set equivalent respectively to cluster is divided into four groups, wherein one group is used to construct prediction model, remaining three groups add It is added to existing model and carries out incremental learning.The input layer number of each model releases the dimension of sexual factor identical i.e. 4 with strong solution, defeated Node layer number is 1 out, that is, exports the single-item power consumption values under the influence of each feature.In addition the population scale that ADE searches for part is set as 50, the number of iterations 100, α, β take 0.8,0.5 respectively.
1) determine that LSTM is counted layer by layer
According to the C that clusters0、C1、C2、C3、C4The single-item energy consumption data for being included constructs the single-item energy consumption prediction mould of such product Type.The other parameters that model is arranged are identical, the LSTM layer of the neural network are gradually increased to 10 layers from 1 layer, record cast is pre- The situation of change of the RMSE of measured value, shown in experimental result Fig. 7.For classification C0With C4Prediction model for, when LSTM layers by 1 When being gradually increased to 3 layers, the value of RMSE is gradually reduced, and when the LSTM number of plies continues growing, RMSE goes out in smaller range It now fluctuates, therefore C0With C4Prediction model can be identified as 3 layers of LSTM structure.And for C1、C2、C3Prediction model for, when When LSTM increases to 2 layers by 1 layer, the value of RMSE is significantly smaller, and when continuing to increase for LSTM layers, RMSE is in smaller range Fluctuation, without obviously less trend, therefore C1、C2、C3Prediction model LSTM layer use two layers.
2) LSTM node layer number is determined
Determination for the middle LSTM node layer number that respectively clusters, the other parameters that each model is arranged are identical, only concept transfer Number obtains the C that respectively clusters by testing0、C1、C2、C3、C4Each node layer number ratio of prediction model be respectively as follows: 4:2:1,2:1, 5:2,5:2,4:3:1, first floor node are gradually increased to 52 by 4, and record cast predicts the variation of error, remaining each node layer number is logical Ratio setting is crossed, experimental results are shown in figure 8.For the C that clusters0Prediction model for, when number of nodes increases to 24 by 4, Each error criterion significantly reduces, when number of nodes increases to 36 by 28, the trend that error criterion is reduced, but only compared with A small range fluctuation, when continuing growing number of nodes, error criterion is on the rise, therefore first layer uses 24 nodes.Together Reason is it is found that C1、C2、C3、C4Prediction model first layer be respectively 16,20,20,20.
2, interpretation of result
Cluster C0、C1、C4There is newly-increased product to be added, therefore to analyze the precision of prediction of newly-increased product single-item energy consumption, comparative test is adopted Clustered corresponding prediction model with this three.By cluster C0、C1、C4Contained historical product energy consumption sample standard deviation according to 3:1 ratio with Machine is divided into training sample and test sample.In the training process, training set equivalent is divided into four groups, and one group for constructing prediction Model, remaining three groups are added to existing model and carry out incremental learning.Increment Density Clustering knot is based on using proposed by the present invention The same LSTM, BP of energy consumption prediction technique, SVR, the DE-LSTM method without incremental learning for closing LSTM carry out incremental learning comparison, The input and output of middle traditional algorithm use the identical input and output of this method, record the precision of prediction and runing time of 10 experiments And averaged, comparing result are as shown in table 6 below.
6 model prediction accuracy of table compares
As seen from the above table, the LSTM wheel hub single-item energy consumption prediction model proposed by the invention based on increment cluster is being predicted Precision and training time are better than other algorithms substantially.From the point of view of precision of prediction, this method is in training stage, test phase and new The RMSE for increasing product energy consumption predicted value is minimum, and the RMSE minimum of training set is down to 0.016524, the RMSE of model training stage 0.012545,0.015129,0.007131 is averagely reduced compared to LSTM, BP, DE-LSTM, is averagely reduced compared to SVR 0.019806, it is seen that the LSTM wheel hub single-item energy consumption prediction model based on increment cluster initially joins model due to using ADE Number optimizes, it is often more important that uses incremental learning strategy, it is contemplated that the timing feature of energy consumption characters, therefore make model Precision of prediction be obviously improved.From the point of view of runing time, the runing time of this method is more than SVR, this is because this method Need to construct the LSTM neural network of multilayer, and the training time and runing time relative to other algorithms are less, this be by Incremental learning can not be carried out in other algorithms, needs re -training therefore to increase runing time when adding training set every time.
Next it is analyzed for the training for adding training data each time, as Fig. 9-11 indicates each institute that clusters When the history energy consumption sample for including is added sequentially to model in four times, each time be added when training precision variation diagram, equally with LSTM, BP, SVR, the DE-LSTM algorithm without incremental learning compare, by the visible algorithm of Fig. 9-11 in first time training The value no significant difference of RMSE, as the addition of incremental data is bright relative to the value reduction of other algorithm RMSE without incremental learning It is aobvious.
Four, conclusion
For solve the problems, such as conventional machines learning model cannot achieve new product unit consumption prediction and can not incremental learning, this hair The bright LSTM wheel hub single-item energy consumption prediction technique for proposing to cluster based on increment, using the dynamic increment Density Clustering based on PCA point The characteristic parameter for analysing wheel hub, to obtain historical product classification similar with newly-increased product, while by Person coefficient and Sexual factor is released in the strong solution that Adaptive-Lasso algorithm filters out single-item energy consumption, and BP algorithm prediction is combined to obtain newly-increased product Sexual factor is released in strong solution, and then is that every a kind of product individually constructs energy consumption prediction model according to cluster result, and it is adaptive poor to use Point evolution algorithm optimizes LSTM, while introducing incremental learning strategy to consider the sequential character of sample, newly-increased when having When product energy consumption sample, incremental update can be carried out to model, save time cost and space consumption.By analysis of experiments, test Demonstrate,proved propose based on increment cluster LSTM wheel hub single-item energy consumption prediction technique validity, make energy consumption predict RMSE down to 0.016524,0.013653 is averagely reduced relative to other algorithms without incremental learning, is realized to the reliable of single-item energy consumption Prediction.Main advantage is as follows:
(1) it is directed to the various problem for causing product classification difficulty of hubless feature parameter complicated and changeable, is proposed based on PCA's Dynamic increment Density Clustering, this method pass through the main feature component of PCA extracting parameter, and the uniqueness of retained product, in turn History single-item classification similar with new product is obtained using clustering algorithm；
(2) aiming at the problem that energy consumption factor importance analysis in wheel hub production process, composite order data, production number Construct energy consumption Factor system according to, crucial consumable component parameter, pass through Pearson coefficient and Adaptive-Lasso algorithm extract it is single Sexual factor is released in the strong solution of product energy consumption, reduces the dimension of prediction input data, and predicts new product using BP neural network Sexual factor value is released in strong solution；
(2) the unpredictable new product energy consumption of conventional machines learning model and can not incremental learning aiming at the problem that, propose to be based on The LSTM incremental update wheel hub unit consumption prediction model of ADE, the model optimize model parameter using ADE algorithm, and will be new The global error of error and historical models between the predicted value and actual value of sample sums it up, and foundation minimizes the error method iteration Model parameter is updated, to guarantee the incremental learning ability of model, keeps model data suitable for production growing number of Actual conditions save space hold and time cost.

Claims (5)

1. a kind of LSTM wheel hub single-item energy consumption prediction technique based on increment cluster, which is characterized in that include the following steps:
(1) the characteristic parameter sample of different model wheel hub, including 12 characteristic parameters: wheel diameter, wheel rim width, center are collected Pitch-row, bolt hole count, pitch diameter, offset distance, weight, spoke number, spoke front pose, center disk moulding, hub material, Manufacturing process；
(2) energy that the order data of collection different model wheel hub, creation data, crucial consumable component parameter, single-item energy consumption data form Consume sample；Wherein order data, creation data, crucial consumable component parameter constitute energy consumption Factor system, and order data includes single Product model, single-item order volume, category sum, creation data includes production time, raw material input amount, operation hours, life Efficiency, rejection rate are produced, is considered in terms of wherein production efficiency adds efficiency and finishing efficiency two by machine, machine adds efficiency unit: part/small When；Finishing efficiency unit: minute/part；Crucial consumable component parameter includes drill bit usage amount, cutter usage amount；
(3) after the character shape parameter of history hubless feature parameter sample in step (1) being carried out labeling processing, using principal component Parameter sample is down to two dimension and obtains data set P by analysis；
(4) cluster operation then is carried out to P using dynamic increment density clustering algorithm, obtains the original result C that clustersi(i=0, 1 ..., k) it ∈ P and peels off and collects O ∈ P；
(5) it when there is newly-increased feature parameter sample, to it using Δ P is obtained after step (3) pretreatment, searches for poly- with original in O ∪ Δ P The reachable data object of cluster density updates cluster result, exports the C that clustersi' (i=0,1 ..., k ')=Ci∪ΔCi, peeling off collects O ' With newly-increased product generic；
(6) sexual factor is released according to the strong solution of energy consumption Factor system analysis single-item energy consumption in step (2), utilizes Pearson system Number rejects the weak related and uncorrelated factor in energy consumption Factor system, carries out followed by Adaptive-Lasso algorithm Second of variables choice obtains the strong explanatory variable of wheel hub single-item energy consumption；
(7) to the strong explanatory variable obtained by step (3) labeling treated history hubless feature parameter, step (6) And single-item power consumption values are standardized；
It (8) is input with the history hubless feature parameter after step (7) standardization, to be standardized by step (7) Treated, and strong explanatory variable is output, constructs the BP prediction model of new wheel hub explanatory variable by force, is able to predict new wheel hub Strong explanatory variate-value；
(9) the single-item energy consumption prediction model that respectively clusters, the history that will respectively cluster energy consumption sample are constructed according to the Historic Clustering result of step (4) This is randomly divided into training set and test set with the ratio of 3:1, and training set is divided into four groups, wherein one group of carry out step (10), remaining set is successively used as a sample increment collection；
(10) the LSTM single-item energy consumption prediction model that respectively clusters is constructed, by strong solution of the training set after step (7) standardization The property released variable is input, and the single-item energy consumption after step (7) standardization is output, real using ADE algorithm in model construction Now to the optimization of LSTM parameter；
(11) in the LSTM single-item energy consumption prediction model that respectively clusters, increment is carried out using sample increment the set pair analysis model in step (9) It updates；
(12) step (11) updated prediction model respectively to cluster is passed through according to the test set test respectively to cluster in step (9), And model evaluation is carried out, it is handled simultaneously for the characteristic parameter sample of product to be predicted by step (3)-(5), and by step Suddenly the BP prediction model of (7)-(8) obtains the strong explanatory variable of new wheel hub, respectively clusters followed by step (11) is updated LSTM single-item energy consumption prediction model is predicted, the predicted value of newly-increased product energy consumption is exported.
2. the LSTM wheel hub single-item energy consumption prediction technique according to claim 1 based on increment cluster, which is characterized in that described Cluster process is carried out to the characteristic parameter sample set P after dimensionality reduction in step (4), is included the following steps:
1) it for the raw data set P after dimensionality reduction, determines the value of coefR, calculates density adjusting parameter σ and density reach distance R, And calculate the density value Density (P of each data object in Pi), obtain Density (P in Pi) maximum local density attracts Sub- Attractor:
In formula,Indicate point PiTo point PjEuclidean distance,The average value of distance between each point in sample, coefR (0 < coefR < 1) be density reach distance original regulation coefficient, n is total sample number；
2) data object in raw data set P is scanned, by density-attractors Attractor and the reachable data object of its density point It is fitted on first cluster C0, and concentrated from initial data and delete the object that clusters；
3) for remaining data collection, another density-attractors Attractor is searchedi, calculate density self-adapting reach distance Radap,i, the density-attractors and the reachable data object of its density are assigned to another C that clustersi, and concentrated from initial data The object that clusters is deleted, and so on, finally the cluster with little data object is put into exceptional value or noise group, at this time To the original result C that clustersi(i=0,1 ..., k) ∈ P, peels off and collects O ∈ P:
Radap,i=α R
α is regulation coefficient in formula, and formula is as follows:
3. the LSTM wheel hub single-item energy consumption prediction technique according to claim 1 based on increment cluster, which is characterized in that described Increment clustering method in step (5) are as follows:
1) for pretreated incremental data set Δ P, Δ P, which is peeled off with original, to be collected O and merges, and is searched in O ∪ Δ P and is respectively clustered with original CiThe reachable data object of density updates cluster result Ci(i=0,1 ..., k) ∈ (P ∪ Δ P), remaining data collection is at this time
2) to remaining data collectionData object carry out clustering, be likely to be present inIn cluster Δ Ci(i=k+1 ..., k ', k ' >=k), update cluster result as C 'i(i=0,1 ..., k ') =Ci(i=0,1 ..., k) ∪ Δ Ci, and peel off and collect O '；
If 3)It then detects whether containing the data object in increment sample Δ P in O ', if containing new product parameter sample This, then compare the new sample point to the middle each point that respectively clusters distance, in order to realize the new product energy consumption prediction, by its elder generation It is tagged to apart from cluster where its closest approach；
4) after for the pretreatment of other incremental data sets, the existing label that clusters collected in O ' that peels off is deleted, step 1) is repeated To 3)；
5) cluster result C ' is exportedi(i=0,1 ..., k ') and peel off set O ' and new product generic.
4. the LSTM wheel hub single-item energy consumption prediction technique according to claim 1 based on increment cluster, which is characterized in that described Single-item energy consumption prediction model construction method in step (10) are as follows:
1) dimension of ADE individual is equal to the weight of LSTM neural network and the summation of threshold number, initializes evolution number G=0, Population scale N, crossover probability f, mutation probability cr, and using root-mean-square error RMSE as fitness function:
GenM is maximum number of iterations in formula, and G is current iteration number, and α, β are the constant in [0.5,1] range, ytIt is true Value, k are the number of data object；
2) fitness value, that is, RMSE of each individual in population is calculated；
3) if minimum RMSE is met the requirements or current iteration number G=GenM, ADE termination iteration in current population, obtain It obtains optimized individual and executes step 5), otherwise continue to execute step 4)；
4) new group is obtained according to adaptive intersection, TSP question and selection operation, G=G+1 is set, is returned to step 3)；
5) optimize the initial connection weight and threshold value for obtaining optimum individual as LSTM based on ADE；And with training sample to LSTM Implement training, and then obtains optimum network.
5. the LSTM wheel hub single-item energy consumption prediction technique according to claim 1 based on increment cluster, which is characterized in that described Sample is increased in step (11) newly to the update method of LSTM prediction model are as follows:
1) for each newly-increased energy consumption sample, first which cluster c judgement belongs toi, corresponding i-th of the LSTM model of i-th of cluster；Belonging to determination After model, model parameter is updated on the basis of model history data training, i.e., is come the parameter after historical data training just The strong explanatory variable of new samples is input in prediction model, obtains new samples by the forward calculation of LSTM by beginningization network Predicted valueBy predicted valueIt is added to the error of actual value y on original error function J (θ)；
In formula, parameter θ=(Wf,Wi,Wc,Wo,bf,bi,bc,bo)
2) model parameter value is updated according to error function J (θ) backpropagation of update:
θ '=(Wf-λ*Δwf,…,Wo-λ*Δwo,bf-λ*Δbf,…,bo-λ*Δbo)
In formula, λ is learning rate, Δ wf、ΔwoWith Δ bf、ΔboBe respectively the weight of neuron and the gradient matrix of offset and to Amount.
CN201910364392.2A 2019-04-30 2019-04-30 LSTM hub single-product energy consumption prediction based on incremental clustering Active CN110070145B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910364392.2A CN110070145B (en) 2019-04-30 2019-04-30 LSTM hub single-product energy consumption prediction based on incremental clustering

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910364392.2A CN110070145B (en) 2019-04-30 2019-04-30 LSTM hub single-product energy consumption prediction based on incremental clustering

Publications (2)

Publication Number Publication Date
CN110070145A true CN110070145A (en) 2019-07-30
CN110070145B CN110070145B (en) 2021-04-27

Family

ID=67369897

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910364392.2A Active CN110070145B (en) 2019-04-30 2019-04-30 LSTM hub single-product energy consumption prediction based on incremental clustering

Country Status (1)

Country Link
CN (1) CN110070145B (en)

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110633844A (en) * 2019-08-25 2019-12-31 天津大学 Building energy system simulation prediction method based on EMD and ANN and application
CN110717628A (en) * 2019-10-09 2020-01-21 浪潮软件股份有限公司 Goods source optimal distribution model construction method, optimal distribution model and optimal distribution method
CN110909818A (en) * 2019-11-30 2020-03-24 航天科技控股集团股份有限公司 TBOX-based dimension reduction system for whole vehicle running data
CN111488924A (en) * 2020-04-07 2020-08-04 中国农业科学院农业信息研究所 Multivariate time sequence data clustering method
CN111832839A (en) * 2020-07-24 2020-10-27 河北工业大学 Energy consumption prediction method based on sufficient incremental learning
CN111881263A (en) * 2020-08-12 2020-11-03 福州大学 Service recommendation online optimization method for intelligent home scene
CN112116140A (en) * 2020-09-10 2020-12-22 同济大学 Building energy consumption prediction method based on twin model
CN112418482A (en) * 2020-10-26 2021-02-26 南京邮电大学 Cloud computing energy consumption prediction method based on time series clustering
CN113155614A (en) * 2021-04-25 2021-07-23 张启志 Concrete compressive strength detection method and system based on similarity determination
CN113780675A (en) * 2021-09-23 2021-12-10 北方健康医疗大数据科技有限公司 Consumption prediction method and device, storage medium and electronic equipment
CN113935557A (en) * 2021-12-21 2022-01-14 中船重工(武汉)凌久高科有限公司 Same-mode energy consumption big data prediction method based on deep learning
CN114202065A (en) * 2022-02-17 2022-03-18 之江实验室 Stream data prediction method and device based on incremental evolution LSTM
CN114417734A (en) * 2022-03-09 2022-04-29 深圳市信润富联数字科技有限公司 Method and device for predicting service life of tool
CN115167279A (en) * 2022-09-07 2022-10-11 中科航迈数控软件(深圳)有限公司 Energy consumption prediction method and system for numerical control machine tool and related equipment
CN116028838A (en) * 2023-01-09 2023-04-28 广东电网有限责任公司 Clustering algorithm-based energy data processing method and device and terminal equipment
CN116823407A (en) * 2023-08-29 2023-09-29 北京国电通网络技术有限公司 Product information pushing method, device, electronic equipment and computer readable medium
CN117113127A (en) * 2023-09-04 2023-11-24 大庆石油管理局有限公司 Oil pumping well energy consumption analysis method based on K-means clustering

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180101313A1 (en) * 2016-10-07 2018-04-12 Prophetstor Data Services, Inc. Method for extending life expectancy of disks in cloud-based service system and system using the same
CN107967542A (en) * 2017-12-21 2018-04-27 国网浙江省电力公司丽水供电公司 A kind of electricity sales amount Forecasting Methodology based on shot and long term memory network
CN108694475A (en) * 2018-06-26 2018-10-23 南京航空航天大学 Short-term time scale photovoltaic cell capable of generating power amount prediction technique based on mixed model
CN108960491A (en) * 2018-06-15 2018-12-07 常州瑞信电子科技有限公司 Method for forecasting photovoltaic power generation quantity based on RBF neural
CN109063416A (en) * 2018-07-23 2018-12-21 太原理工大学 Gene expression prediction technique based on LSTM Recognition with Recurrent Neural Network
CN109396576A (en) * 2018-09-29 2019-03-01 郑州轻工业学院 Stability of EDM and power consumption state Optimal Decision-making system and decision-making technique based on deep learning
CN109523386A (en) * 2018-10-18 2019-03-26 广东工业大学 A kind of investment portfolio risk prediction technique of GMM in conjunction with LSTM

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180101313A1 (en) * 2016-10-07 2018-04-12 Prophetstor Data Services, Inc. Method for extending life expectancy of disks in cloud-based service system and system using the same
CN107967542A (en) * 2017-12-21 2018-04-27 国网浙江省电力公司丽水供电公司 A kind of electricity sales amount Forecasting Methodology based on shot and long term memory network
CN108960491A (en) * 2018-06-15 2018-12-07 常州瑞信电子科技有限公司 Method for forecasting photovoltaic power generation quantity based on RBF neural
CN108694475A (en) * 2018-06-26 2018-10-23 南京航空航天大学 Short-term time scale photovoltaic cell capable of generating power amount prediction technique based on mixed model
CN109063416A (en) * 2018-07-23 2018-12-21 太原理工大学 Gene expression prediction technique based on LSTM Recognition with Recurrent Neural Network
CN109396576A (en) * 2018-09-29 2019-03-01 郑州轻工业学院 Stability of EDM and power consumption state Optimal Decision-making system and decision-making technique based on deep learning
CN109523386A (en) * 2018-10-18 2019-03-26 广东工业大学 A kind of investment portfolio risk prediction technique of GMM in conjunction with LSTM

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
GANG XU等: "Short- Term Prediction of Wind Power Based on Adaptive LSTM", 《2018 2ND IEEE CONFERENCE ON ENERGY INTERNET AND ENERGY SYSTEM INTEGRATION (EI2)》 *
SUMIT KUMAR等: "Energy Load Forecasting using Deep Learning Approach-LSTM and GRU in Spark Cluster", 《2018 FIFTH INTERNATIONAL CONFERENCE ON EMERGING APPLICATIONS OF INFORMATION TECHNOLOGY (EAIT)》 *
孙秋野等: "基于 GAN 技术的自能源混合建模与参数辨识方法", 《自动化学报》 *
陈佳等: "一种基于LSTM神经网络的短期用电负荷预测方法", 《电力信息与通信技术》 *
陈周林等: "改进PSO-BP网络预测模型在造纸能耗预测中的应用", 《轻工科技》 *

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110633844A (en) * 2019-08-25 2019-12-31 天津大学 Building energy system simulation prediction method based on EMD and ANN and application
CN110633844B (en) * 2019-08-25 2023-02-24 天津大学 Building energy system simulation prediction method based on EMD and ANN and application
CN110717628A (en) * 2019-10-09 2020-01-21 浪潮软件股份有限公司 Goods source optimal distribution model construction method, optimal distribution model and optimal distribution method
CN110717628B (en) * 2019-10-09 2023-05-23 浪潮软件股份有限公司 Goods source optimizing model construction method, optimizing model and optimizing method
CN110909818B (en) * 2019-11-30 2022-08-09 航天科技控股集团股份有限公司 TBOX-based dimension reduction system for whole vehicle running data
CN110909818A (en) * 2019-11-30 2020-03-24 航天科技控股集团股份有限公司 TBOX-based dimension reduction system for whole vehicle running data
CN111488924A (en) * 2020-04-07 2020-08-04 中国农业科学院农业信息研究所 Multivariate time sequence data clustering method
CN111488924B (en) * 2020-04-07 2024-04-26 中国农业科学院农业信息研究所 Multivariable time sequence data clustering method
CN111832839A (en) * 2020-07-24 2020-10-27 河北工业大学 Energy consumption prediction method based on sufficient incremental learning
CN111881263A (en) * 2020-08-12 2020-11-03 福州大学 Service recommendation online optimization method for intelligent home scene
CN112116140A (en) * 2020-09-10 2020-12-22 同济大学 Building energy consumption prediction method based on twin model
CN112116140B (en) * 2020-09-10 2022-09-20 同济大学 Building energy consumption prediction method based on twin model
CN112418482A (en) * 2020-10-26 2021-02-26 南京邮电大学 Cloud computing energy consumption prediction method based on time series clustering
CN112418482B (en) * 2020-10-26 2022-08-26 南京邮电大学 Cloud computing energy consumption prediction method based on time series clustering
CN113155614B (en) * 2021-04-25 2023-07-28 张启志 Concrete compressive strength detection method and system based on similarity judgment
CN113155614A (en) * 2021-04-25 2021-07-23 张启志 Concrete compressive strength detection method and system based on similarity determination
CN113780675B (en) * 2021-09-23 2024-01-09 北方健康医疗大数据科技有限公司 Consumption prediction method and device, storage medium and electronic equipment
CN113780675A (en) * 2021-09-23 2021-12-10 北方健康医疗大数据科技有限公司 Consumption prediction method and device, storage medium and electronic equipment
CN113935557A (en) * 2021-12-21 2022-01-14 中船重工(武汉)凌久高科有限公司 Same-mode energy consumption big data prediction method based on deep learning
CN114202065A (en) * 2022-02-17 2022-03-18 之江实验室 Stream data prediction method and device based on incremental evolution LSTM
CN114417734B (en) * 2022-03-09 2022-07-12 深圳市信润富联数字科技有限公司 Method and device for predicting service life of tool
CN114417734A (en) * 2022-03-09 2022-04-29 深圳市信润富联数字科技有限公司 Method and device for predicting service life of tool
CN115167279A (en) * 2022-09-07 2022-10-11 中科航迈数控软件(深圳)有限公司 Energy consumption prediction method and system for numerical control machine tool and related equipment
CN116028838A (en) * 2023-01-09 2023-04-28 广东电网有限责任公司 Clustering algorithm-based energy data processing method and device and terminal equipment
CN116028838B (en) * 2023-01-09 2023-09-19 广东电网有限责任公司 Clustering algorithm-based energy data processing method and device and terminal equipment
CN116823407A (en) * 2023-08-29 2023-09-29 北京国电通网络技术有限公司 Product information pushing method, device, electronic equipment and computer readable medium
CN116823407B (en) * 2023-08-29 2024-01-12 北京国电通网络技术有限公司 Product information pushing method, device, electronic equipment and computer readable medium
CN117113127A (en) * 2023-09-04 2023-11-24 大庆石油管理局有限公司 Oil pumping well energy consumption analysis method based on K-means clustering
CN117113127B (en) * 2023-09-04 2024-03-29 大庆石油管理局有限公司 Oil pumping well energy consumption analysis method based on K-means clustering

Also Published As

Publication number Publication date
CN110070145B (en) 2021-04-27

Similar Documents

Publication Publication Date Title
CN110070145A (en) LSTM wheel hub single-item energy consumption prediction based on increment cluster
CN109754113B (en) Load prediction method based on dynamic time warping and long-and-short time memory
CN109685252B (en) Building energy consumption prediction method based on cyclic neural network and multi-task learning model
KR102139358B1 (en) Process control method using machine learning-based platform, computer program medium for performing same, and process control appratus
CN112990556A (en) User power consumption prediction method based on Prophet-LSTM model
CN111191835B (en) IES incomplete data load prediction method and system based on C-GAN migration learning
CN110083125B (en) Machine tool thermal error modeling method based on deep learning
CN106056127A (en) GPR (gaussian process regression) online soft measurement method with model updating
CN110571792A (en) Analysis and evaluation method and system for operation state of power grid regulation and control system
CN109255726A (en) A kind of ultra-short term wind power prediction method of Hybrid Intelligent Technology
CN109886465A (en) A kind of distribution network load prediction technique based on intelligent electric meter user&#39;s clustering
CN114757104B (en) Method for constructing hydraulic real-time regulation and control model of series gate group water transfer project
CN111915092A (en) Ultra-short-term wind power prediction method based on long-time and short-time memory neural network
CN113947182B (en) Traffic flow prediction model construction method based on dual-stage stacked graph convolution network
CN111832839B (en) Energy consumption prediction method based on sufficient incremental learning
CN114492191A (en) Heat station equipment residual life evaluation method based on DBN-SVR
CN112330052A (en) Distribution transformer load prediction method
CN112884236A (en) Short-term load prediction method and system based on VDM decomposition and LSTM improvement
CN116842354A (en) Feature selection method based on quantum artificial jellyfish search mechanism
CN113762591B (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning
CN109408896B (en) Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production
CN114596726A (en) Parking position prediction method based on interpretable space-time attention mechanism
Yang Combination forecast of economic chaos based on improved genetic algorithm
CN117439053A (en) Method, device and storage medium for predicting electric quantity of Stacking integrated model
CN115481788B (en) Phase change energy storage system load prediction method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CP02 Change in the address of a patent holder

Address after: 300457 room 803, Ronghui building, No. 58, Dongting Road, economic and Technological Development Zone, Binhai New Area, Tianjin

Patentee after: Tianjin Development Zone Jingnuo Hanhai Data Technology Co.,Ltd.

Address before: 300401 door 9, building 1, Science Park, Hebei University of technology, Shuangkou Town, Beichen District, Tianjin

Patentee before: Tianjin Development Zone Jingnuo Hanhai Data Technology Co.,Ltd.

CP02 Change in the address of a patent holder