CN110070145B - LSTM hub single-product energy consumption prediction based on incremental clustering - Google Patents

LSTM hub single-product energy consumption prediction based on incremental clustering Download PDF

Info

Publication number
CN110070145B
CN110070145B CN201910364392.2A CN201910364392A CN110070145B CN 110070145 B CN110070145 B CN 110070145B CN 201910364392 A CN201910364392 A CN 201910364392A CN 110070145 B CN110070145 B CN 110070145B
Authority
CN
China
Prior art keywords
energy consumption
product
cluster
lstm
data
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.)
Active
Application number
CN201910364392.2A
Other languages
Chinese (zh)
Other versions
CN110070145A (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 Hanhai Data Technology Co ltd
Original Assignee
Tianjin Development Zone Jingnuo Hanhai 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 Hanhai Data Technology Co ltd filed Critical Tianjin Development Zone Jingnuo Hanhai 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

Images

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

Abstract

The invention discloses LSTM hub single-product energy consumption prediction based on incremental clustering, and relates to the technical field of hub single-product energy consumption prediction. According to the invention, dynamic incremental density clustering based on PCA is adopted to realize clustering analysis of hub characteristic parameters, historical product categories to which new products belong are obtained, Pearson coefficients and Adaptive-Lasso algorithms are utilized to screen out single product energy consumption strong explanatory factors based on an energy consumption influence factor system, prediction of new product strong explanatory variables is realized by BP, an LSTM energy consumption prediction model is built for each cluster product to realize effective prediction of new product unit consumption, optimization of LSTM is realized by ADE, and an incremental learning strategy is introduced to realize dynamic update of the model. The invention verifies the effectiveness of the prediction method, the Root Mean Square Error (RMSE) of energy consumption prediction is reduced to 0.016524, compared with other algorithms without incremental learning, the method averagely reduces 0.013653, the ADE search performance is better, the RMSE of a training set is averagely reduced by 0.004089 compared with that of DE-LSTM with incremental learning, and the running time is effectively shortened.

Description

LSTM hub single-product energy consumption prediction based on incremental clustering
Technical Field
The invention relates to the technical field of single-product energy consumption prediction of hubs, in particular to an LSTM (least squares metric) hub single-product energy consumption prediction method based on incremental clustering.
Background
The rapid development of the hub industry in China is driven under the vigorous demand of the whole automobile market, more than 300 hub manufacturers exist in China at present, the two-digit increase speed of the automobile hub yield is kept, the demand of hubs in the after-sale market of automobiles in 2022 in China is estimated to be about 604 thousands, and the hub products enter the 'China manufacturing era'. Therefore, in order to respond to the market demand and meet the individual customization requirements of customers, hub manufacturers gradually develop a multi-class small-batch production mode. However, when a new product without production experience needs to be produced, because historical energy consumption data are not accumulated, the traditional model cannot predict the energy consumption of the new product, and the energy consumption directly influences the cost of the single product of the hub, so that the research on the energy consumption of the new product of the hub is of great significance.
In the aspect of establishing an energy consumption prediction model, due to the deep fusion of intelligent production and internet of things technology, data acquisition in the industrial production process is convenient and fast, various intelligent prediction technologies are used for energy demand management to accurately predict future energy demands, and time series prediction, gray prediction, support vector machines, neural network prediction and the like are mainly adopted at present, so that better research results are obtained. An article [ Gunn macro, et al, an air transportation enterprise energy consumption prediction model (English) based on wavelet-ARIMA [ J ] machine tool and hydraulic pressure, 2018,46(06):13-17+42 ] proposes an ARIMA prediction model based on two-scale wavelet decomposition; an article [ Suvqin. prediction simulation of energy consumption of a paper making enterprise process [ J ] computer simulation, 2016,33(08):438 + 442+447 ] proposes an ARIMA model based on Bayes to predict energy consumption by improving a maximum likelihood parameter estimation method in the ARIMA model. An article [ Lirisxin and the like ] energy consumption prediction based on a gray Markov model [ J ] Chinese scientific and technical information, 2018(15):74-75 ] proposes a view of introducing a Markov chain on the basis of the gray prediction model aiming at the prediction of the total energy consumption of China; an article [ Liujia, etc.. an airline energy consumption prediction (English) based on a metabolism gray Markov-ARMA model [ J ] machine tool and hydraulic pressure, 2017,45(18):55-62 ] corrects a prediction result of GM (1,1) by the Markov, eliminates old data losing timeliness in the model by a metabolism method, and corrects residual errors by a sliding time window and the ARMA model; an article [ Durui Zhi, and the like, application of a GM-WLSSVM model in the prediction of the electric energy consumption of the office building [ J ] computer application and software, 2018,35(09):44-49+55 ] uses a gray model to select different samples to perform multi-type prediction on the same time interval, and then uses a weighted least square support vector machine model to combine prediction results, thereby realizing the short-term prediction of the electric energy of the office building; an article [ Wangkun et al, LSSVM for airport energy consumption prediction [ J ] based on EMD and fruit fly parameter optimization [ computer age, 2017(04):35-40 ] proposes an energy consumption prediction method of a least square support vector machine combining empirical mode decomposition and fruit fly parameter optimization; an article [ summer Weijun and the like, research on an energy consumption prediction model of a paper making enterprise based on PSO-LSSVR [ J ] computer measurement and control, 2013,21(12): 3433-; an improved grey depth belief network combined prediction model is provided by an article [ Guo Xiao Silent, et al, civil aviation airport electricity short-term energy consumption optimized prediction simulation [ J ] computer simulation, 2018,35(09):31-36 ], so that the prediction accuracy of the civil aviation airport electricity short-term energy consumption model is improved; an article [ Chenzhonglin et al, application of an improved PSO-BP network prediction model [ J/OL ]. light engineering technology, 2018(11):91-94[2019-01-08 ] ] in papermaking energy consumption prediction establishes a BP neural network energy consumption prediction model based on improved particle swarm optimization; an article [ the cheering tablet and the like ] operation energy consumption prediction [ J ] of an HVAC system of a green office building based on ANN, building energy conservation, 2017,45(10):1-5 ] establishes a neural network prediction model of a classified multilayer perceptron, and realizes energy consumption prediction of a heating ventilation air-conditioning system in the office building; an LSTM neural network-based short-term power load prediction method [ J ] power information and communication technology, 2017,15(09):19-25 ] uses power load data as training data and an output label, and an LSTM-based power load prediction model is established by an iterative training method.
As can be seen from the above documents, the prediction problem of analyzing the energy consumption of a single product is not researched, and the hybrid prediction model can improve the overall performance of the model better. Because the relation between the energy consumption and the influence factors is nonlinear, the neural network can be used for realizing accurate prediction, but the time modeling capability of the traditional feedforward neural network is quite limited, and the LSTM can solve the problem of learning long-term dependence under the condition that the output prediction value depends on the long history of the input feature sequence. However, how to obtain the strong interpretative influence factors of the energy consumption is the research difficulty and the foundation of constructing a prediction model, and data is continuously increased along with the time lapse.
Disclosure of Invention
The invention aims to provide an LSTM hub single-product energy consumption prediction method based on incremental clustering. The method aims to adopt PCA to obtain a historical single product category similar to a new product after dimension reduction of characteristic parameters for determining a historical hub production mode; secondly, an energy consumption influence factor system is constructed based on order data, production data and key consumable part parameters, strong interpretability variables strongly related to the energy consumption of the single product are obtained by utilizing the Pearson coefficient and Adaptive-Lasso, and a strong interpretability factor value of the new product is predicted by utilizing a BP neural network; and finally, constructing a single-product energy consumption prediction model of each cluster hub on the basis of the method, and providing an LSTM increment updating hub single-product energy consumption prediction model based on ADE (adaptive data analysis), wherein the model searches initial parameters of the LSTM model by adopting an ADE algorithm, and updates the prediction model when a new sample is added. The method realizes effective prediction of single-product energy consumption of the new hub and incremental updating of the model, improves prediction precision, and reduces data storage space and model calculation time.
The technical scheme adopted by the invention is an LSTM hub single-product energy consumption prediction method based on incremental clustering, which is characterized by comprising the following steps:
(1) collecting characteristic parameter samples of hubs of different models, wherein the characteristic parameter samples comprise 12 characteristic parameters: the wheel rim comprises a wheel rim diameter, a wheel rim width, a center hole distance, a bolt hole number, a pitch circle diameter, an offset distance, weight, a wheel spoke number, a wheel spoke front face model, a center disc surface model, a wheel hub material and a manufacturing process;
(2) collecting energy consumption samples consisting of order data, production data, key consumable part parameters and single-product energy consumption data of hubs of different models; the system comprises an order data acquisition system, a data processing system and a data processing system, wherein the order data, the production data and key consumable part parameters form an energy consumption influence factor system, the order data comprises a unit type number, a unit order amount and a unit type total number, the production data comprises production time, raw material input amount, equipment operation time, production efficiency and rejection rate, the production efficiency is considered by two aspects of machining efficiency (parts/hour) and finishing efficiency (minutes/parts), and the key consumable part parameters comprise drill bit usage and cutter tool usage;
(3) performing labeling processing on character type parameters of the historical hub characteristic parameter sample in the step (1), and reducing the parameter sample to two dimensions by adopting principal component analysis to obtain a data set P;
(4) then, clustering operation is carried out on the P by using a dynamic incremental density clustering algorithm to obtain an original clustering result Ci(i ═ 0,1, …, k) e P and an outlier O e P;
(5) when a new characteristic parameter sample is added, preprocessing the sample by the step (3) to obtain delta P, searching a data object which can reach the original clustering density in the O U delta P, updating a clustering result, and outputting a cluster C'i(i＝0,1,…,k′)＝Ci∪ΔCiThe outlier O' and the category to which the newly added product belongs;
(6) analyzing strong explanatory factors of the energy consumption of the single product according to the energy consumption influence factor system in the step (2), eliminating weak relevant and irrelevant factors in the energy consumption influence factor system by using a Pearson coefficient, and then performing secondary variable selection by using an Adaptive-Lasso algorithm to obtain strong explanatory variables of the energy consumption of the single product of the hub;
(7) standardizing the historical hub characteristic parameters subjected to labeling processing in the step (3), the strong interpretable variable obtained in the step (6) and the single product energy consumption value;
(8) taking the historical hub characteristic parameters subjected to the standardization processing in the step (7) as input, taking the strong interpretable variable subjected to the standardization processing in the step (7) as output, and constructing a BP prediction model of the strong interpretable variable of the new hub so as to predict the strong interpretable variable value of the new hub;
(9) constructing a single-product energy consumption prediction model of each cluster according to the historical clustering result in the step (4), randomly dividing each historical energy consumption sample of each cluster into a training set and a test set according to the ratio of 3:1, dividing the training set into four groups, performing the step (9) on one group, and sequentially using the rest groups as a sample increment set;
(10) constructing a clustered LSTM single-product energy consumption prediction model, taking the strong interpretability variable of the training set subjected to the standardization in the step (7) as input, taking the single-product energy consumption subjected to the standardization in the step (7) as output, and realizing the optimization of LSTM parameters by using an ADE algorithm in the model construction;
(11) in each cluster LSTM single-product energy consumption prediction model, carrying out incremental updating on the model by using the sample increment set in the step (9);
(12) and (3) testing the prediction model of each cluster updated in the step (11) according to the test set of each cluster in the step (9), evaluating the model, processing the characteristic parameter samples of the product to be predicted in the steps (3) to (5) and obtaining the strong interpretable variable of the new hub through the BP prediction model in the steps (7) to (8), predicting by using the LSTM single-product energy consumption prediction model of each cluster updated in the step (11), and outputting the predicted value of the energy consumption of the newly-increased product.
A further technical scheme is that, in the step (4), the clustering process is performed on the feature parameter sample set P after the dimension reduction, and the method includes the following steps:
1) for the original data set P after dimensionality reduction, determining the value of coefR, calculating a Density adjustment parameter sigma and a Density reachable distance R, and calculating the Density value sensitivity (P) of each data object in Pi) To obtain the sensitivity (P) in Pi) Maximum local density Attractor:
Figure GDA0002950232510000051
Figure GDA0002950232510000061
Figure GDA0002950232510000062
in the formula (I), the compound is shown in the specification,
Figure GDA0002950232510000063
representing point PiTo point PjThe Euclidean distance of (a) is,
Figure GDA0002950232510000064
coefR (0), an average of the distances between points in the sample<coefR<1) The original adjustment coefficient of the density reachable distance is obtained, and n is the total number of samples;
2) scanning data objects in an original data set P, and allocating a density Attractor and data objects with reachable density to a first cluster C0And deleting the cluster object from the original data set;
3) for the remaining dataset, look up another density AttractoriCalculating the adaptive density reachable distance Radap,iAssign the density attractor and the data object whose density is reachable to another cluster CiDeleting the cluster object from the original data set, repeating the above steps, and finally putting the cluster with less data objects into an abnormal value or noise group, and obtaining an original cluster result Ci(i ═ 0,1, …, k) ∈ P, outlier O ∈ P:
Radap,i＝αR
where α is the adjustment factor, the formula is as follows:
Figure GDA0002950232510000065
the further technical scheme is that the incremental clustering method in the step (5) comprises the following steps:
1) for the preprocessed incremental data set delta P, combining the delta P with the original outlier O, and searching each cluster C in the O U delta PiUpdating clustering result C for data objects with reachable densityi(i ═ 0,1, …, k) ∈ (P ≧ Δ P) when the remaining dataset is
Figure GDA0002950232510000066
2) For the remaining data sets
Figure GDA0002950232510000067
The data object is subjected to clustering analysis to obtain the data object possibly existing in
Figure GDA0002950232510000068
Cluster Δ C in (1)i(i ═ k +1, …, k ', k ' ≧ k), and the update clustering result is C 'i(i＝0,1,…,k′)＝Ci(i＝0,1,…,k)∪ΔCiAnd an outlier O';
3) if it is
Figure GDA0002950232510000069
Detecting whether the O 'contains a data object in the incremental sample delta P, if the O' contains a new product parameter sample, comparing the distance from the new sample point to each point in each cluster, and marking the new sample point to the cluster where the closest point is located in order to realize energy consumption prediction of the new product;
4) after other incremental data sets are preprocessed, deleting the clustering mark in the existing cluster O', and repeatedly executing the steps 1) to 3);
5) outputting a clustering result C'i(i ═ 0,1, …, k ') and the outlier O', and the category to which the new product belongs.
The further technical scheme is that the method for constructing the single-product energy consumption prediction model in the step (10) comprises the following steps:
1) the dimension of the ADE individual is equal to the sum of the weight of the LSTM neural network and the threshold number, the initial evolution time G is 0, the population size N, the cross probability f and the variation probability cr, and the root mean square error RMSE is adopted as a fitness function:
Figure GDA0002950232510000071
Figure GDA0002950232510000072
Figure GDA0002950232510000073
in the formula, GenM is the maximum iteration number, G is the current iteration number, and alpha and beta are [0.5,1]Constant within the range, ytK is the true value, and k is the number of data objects;
2) calculating the fitness value (RMSE) of each individual in the population;
3) if the minimum RMSE in the current population meets the requirement or the current iteration time G is GenM, the ADE terminates the iteration, obtains the best individual and executes the step 5), otherwise, the step 4) is continuously executed;
4) obtaining a new group according to the self-adaptive intersection, the self-adaptive variation and the selection operation, setting G as G +1, and returning to execute the step 3);
5) obtaining optimal individuals as initial connection weights and thresholds of the LSTM based on ADE optimization; and training the LSTM by using the training samples to further obtain the optimal network.
The further technical scheme is that the updating method of the LSTM prediction model by adding the samples in the step (11) comprises the following steps:
1) for each newly-increased energy consumption sample, firstly, judging which cluster c belongs toiThe ith cluster corresponds to the ith LSTM model; after the model is determined, the model parameters are updated on the basis of the historical data training of the model, namely, the parameters after the historical data training are used for initializing the network, the strong explanatory variable of the new sample is input into the prediction model, and the predicted value of the new sample is obtained through the forward calculation of the LSTM
Figure GDA0002950232510000081
Will predict the value
Figure GDA0002950232510000082
The error from the actual value y is added to the original error function J (theta)) The above step (1);
Figure GDA0002950232510000083
wherein the parameter θ ═ Wf,Wi,Wc,Wo,bf,bi,bc,bo)
2) Updating the model parameter values according to the updated error function J (theta) back propagation:
θ′＝(Wf-λ*Δwf,…,Wo-λ*Δwo,bf-λ*Δbf,…,bo-λ*Δbo)
wherein λ is a learning rate, Δ wf、ΔwoAnd Δ bf、ΔboGradient matrices and vectors of weights and offsets of the neurons, respectively.
Adopt the produced beneficial effect of above-mentioned technical scheme to lie in:
according to the LSTM hub single-product energy consumption prediction method based on incremental clustering, the dynamic incremental density clustering based on PCA is adopted to realize the clustering analysis of hub characteristic parameters, so that the historical product category of a newly added product is obtained, on the basis of an energy consumption influence factor system, Pearson coefficients and Adaptive-Lasso algorithms are utilized to screen out single-product energy consumption strong explanatory factors, further, BP is utilized to realize the prediction of new product strong explanatory variables, finally, an LSTM energy consumption prediction model is constructed for each cluster product to realize the effective prediction of new product single consumption, ADE is utilized to realize the optimization of LSTM, and an incremental learning strategy is introduced to realize the dynamic update of the model. Through experimental analysis, the effectiveness of the proposed LSTM hub single-product energy consumption prediction method based on incremental clustering is verified, the root mean square error RMSE of energy consumption prediction is reduced to 0.016524, compared with other algorithms without incremental learning, the method is averagely reduced by 0.013653, meanwhile, the ADE searching performance is better, the RMSE of a training set is averagely reduced by 0.004089 compared with that of DE-LSTM with incremental learning, and the running time is effectively shortened.
Drawings
The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
FIG. 1 is a structural diagram of an LSTM hub single product energy consumption prediction method based on incremental clustering;
FIG. 2 is a flow chart of dynamic incremental density clustering based on PCA of the present invention;
FIG. 3 is a flowchart of the ADE based LSTM incremental update hub specific consumption prediction of the present invention;
FIG. 4 is a graph of the cluster analysis results of the present invention;
fig. 5 and 6 are RMSE change diagrams of the number of hidden layer nodes and X8 and the number of hidden layer nodes and X9, respectively;
FIG. 7 is a graph of the number of LSTM layers versus the RMSE variation for each cluster prediction model;
FIG. 8 is a graph of the number of LSTM layer nodes versus the RMSE variation for each cluster prediction model;
FIG. 9 shows a cluster C0The energy consumption increment data are compared with the training precision of the five algorithms;
FIG. 10 shows a cluster C1The energy consumption increment data are compared with the training precision of the five algorithms;
FIG. 11 shows a cluster C4The energy consumption increment data are compared with the training precision of the five algorithms;
Detailed Description
The invention provides a dynamic incremental density clustering algorithm based on PCA, namely, a historical single product category similar to a new product is obtained by utilizing the clustering algorithm based on characteristic parameter data which symbolizes a hub production mode; secondly, analyzing strong explanatory factors of single product energy consumption by utilizing a Pearson coefficient and an Adaptive-Lasso algorithm, and predicting a strong explanatory factor value of a new product by utilizing a BP neural network; finally, an LSTM increment updating hub unit consumption prediction model based on the ADE is provided, the influence of initialization parameters on model precision is weakened by the model through the ADE algorithm, an increment learning strategy is introduced, and dynamic updating of the model is achieved.
First, the theoretical basis of the method of the invention
1. Principal Component Analysis (PCA): transforming raw data into a set of linearly independent (unit orthogonal) representations of each dimension through linear transformation can be used to extract the main characteristic components of the data while preserving as much variability as possible, which is often used for dimensionality reduction of high-dimensional data.
2. Density-based dynamic incremental clustering: defining the density attractor with the maximum data point density, searching local density attractors and sample points with the reachable density from the incomplete clustered residual data set each time, thereby clustering the attractors into a class, and adaptively adjusting the reachable density distance in order to be suitable for a dynamic incremental database.
3. Pearson correlation coefficient (Pearson coefficient): the method is used for reflecting the degree of correlation between two variables, and can be used for calculating the similarity between the features and the categories in machine learning, namely judging whether the extracted features and the categories are positively correlated, negatively correlated or have no degree of correlation.
4. Adaptive-Lasso algorithm: the method is an improvement on the Lasso algorithm, a more refined model is obtained by constructing a penalty function, so that the model compresses a plurality of coefficients, and the coefficients are set to be zero, thereby realizing the purpose of quickly and effectively extracting important variables when the variables are numerous, and simplifying the model.
5. Long-Short Term Memory neural network (LSTM): the RNN is a variant of the RNN, namely on the basis of a common RNN, memory units are added in each nerve unit of a hidden layer, so that memory information on a time sequence is controllable, and the memory and forgetting degree of previous information and current information can be controlled through a plurality of controllable gates (a forgetting gate, an input gate, a candidate gate and an output gate) when the memory units are transmitted among the units of the hidden layer every time, so that the RNN has a long-term memory function.
6. The self-adaptive differential evolution algorithm: (Adaptive differential evolution algorithm, ADE): the improvement of the standard differential evolution algorithm is realized by adaptively adjusting the scaling factor and the cross probability in the calculation process.
Secondly, designing an LSTM hub single-product energy consumption prediction method based on incremental clustering:
the LSTM hub single-product energy consumption prediction method based on incremental clustering has the following working principle: when single-product energy consumption prediction needs to be carried out on a newly added product, firstly, PCA is adopted to reduce characteristic parameters for determining a historical hub production mode to two dimensions, and then a dynamic incremental density clustering algorithm is utilized to obtain a historical single-product category similar to a new product; secondly, an energy consumption influence factor system is built based on order data, production data and key consumable part parameters, strong interpretable variables strongly related to the energy consumption of the single product are obtained by utilizing a Pearson coefficient and an Adaptive-Lasso algorithm, the characteristic parameters of the hub are used as input, the strong interpretable variables are used as output, and a BP prediction model of the new strong interpretable variables is built; and finally, constructing a single-product energy consumption prediction model of each cluster hub on the basis of the method, weakening the influence of initialization parameters on the model by using an ADE algorithm, considering the problem of processing a newly-added energy consumption sample, introducing an incremental learning strategy, determining a prediction model to be updated by judging the cluster to which the new sample belongs, taking the historical parameters of the determined model as the initial parameters of the model, adding the error between the predicted value and the actual value of the new sample into the overall error, and iteratively updating the model parameters by using an error minimization method to ensure that the model can continuously process the added newly-added sample.
1. LSTM hub single-product energy consumption prediction method structure based on incremental clustering
According to the invention, the historical product category similar to the production mode of a newly added product is obtained through dynamic incremental density clustering based on PCA, and an LSTM incremental updating hub unit consumption prediction model based on ADE is constructed for each cluster based on the clustering result, so that the unit consumption prediction of the unit product is realized. As shown in fig. 1, when a new product characteristic parameter sample exists, the new product characteristic parameter sample is reduced to two dimensions by PCA, a clustering result is updated by a dynamic incremental density clustering algorithm to obtain a historical product category to which the new product belongs, meanwhile, screening of energy consumption strong explanatory variables is realized by Pearson coefficients and Adaptive-Lasso algorithm, a strong explanatory factor of the new product is obtained by BP neural network prediction, and then the strong explanatory factor is input to a corresponding ADE optimized LSTM model to obtain a unit energy consumption value of the new product, and when the new product characteristic parameter sample exists, a model error function is updated according to an error between a predicted value and a true value of the new product, model parameters are adjusted, and incremental updating of the model is realized.
2. Specific implementation of the Algorithm
The LSTM hub single-product energy consumption prediction method based on incremental clustering comprises the following steps:
(1) collecting characteristic parameter samples of hubs of different models, wherein the characteristic parameter samples comprise 12 characteristic parameters: the wheel rim is characterized by comprising the following steps of (1) rim diameter (p1), rim width (p2), center hole distance (p3), bolt hole number (p4), pitch circle diameter (p5), offset distance (p6), weight (p7), spoke number (p8), spoke front modeling (p9), center disc surface modeling (p10), hub material (p11) and manufacturing process (p 12);
(2) collecting energy consumption samples consisting of order data, production data, key consumable part parameters and single-product energy consumption data of hubs of different models; the order data, the production data and the key consumable part parameters form an energy consumption influence factor system, the order data comprises a single product model (X1), a single product order amount (X2) and a total number of product types (X3), the production data comprises production time (X4), raw material input amount (X5), equipment running time (X6), production efficiency (machining efficiency X7, finishing efficiency X8) and rejection rate (X9), and the key consumable part parameters comprise drill bit usage amount (X10) and cutter usage amount (X11);
(3) performing labeling processing on character type parameters of the historical hub characteristic parameter sample in the step (1), and reducing the parameter sample to two dimensions by adopting principal component analysis to obtain a data set P;
(4) then, clustering operation is carried out on the P by using a dynamic incremental density clustering algorithm to obtain an original clustering result Ci(i ═ 0,1, …, k) e P and an outlier O e P;
(5) when a new characteristic parameter sample is added, preprocessing the sample by the step (3) to obtain delta P, searching a data object which can reach the original clustering density in the O U delta P, updating a clustering result, and outputting a cluster C'i(i＝0,1,…,k′)＝Ci∪ΔCiThe outlier O' and the category to which the newly added product belongs;
(6) analyzing strong explanatory factors of the energy consumption of the single product according to the energy consumption influence factor system in the step (2), eliminating weak relevant and irrelevant factors in the energy consumption influence factor system by using a Pearson coefficient, and then performing secondary variable selection by using an Adaptive-Lasso algorithm to obtain strong explanatory variables of the energy consumption of the single product of the hub;
(7) standardizing the historical hub characteristic parameters subjected to labeling processing in the step (3), the strong interpretable variable obtained in the step (6) and the single product energy consumption value;
(8) taking the historical hub characteristic parameters subjected to the standardization processing in the step (7) as input, taking the strong interpretable variable subjected to the standardization processing in the step (7) as output, and constructing a BP prediction model of the strong interpretable variable of the new hub so as to predict the strong interpretable variable value of the new hub;
(9) constructing a single-product energy consumption prediction model of each cluster according to the historical clustering result in the step (4), randomly dividing each historical energy consumption sample of each cluster into a training set and a test set according to the ratio of 3:1, dividing the training set into four groups, performing the step (9) on one group, and sequentially using the rest groups as a sample increment set;
(10) constructing a clustered LSTM single-product energy consumption prediction model, taking the strong interpretability variable of the training set subjected to the standardization in the step (7) as input, taking the single-product energy consumption subjected to the standardization in the step (7) as output, and realizing the optimization of LSTM parameters by using an ADE algorithm in the model construction;
(11) in each cluster LSTM single-product energy consumption prediction model, carrying out incremental updating on the model by using the sample increment set in the step (9);
(12) and (3) testing the prediction model of each cluster updated in the step (11) according to the test set of each cluster in the step (9), evaluating the model, processing the characteristic parameter samples of the product to be predicted in the steps (3) to (5) and obtaining the strong interpretable variable of the new hub through the BP prediction model in the steps (7) to (8), predicting by using the LSTM single-product energy consumption prediction model of each cluster updated in the step (11), and outputting the predicted value of the energy consumption of the newly-increased product.
In the embodiment of the present invention, the clustering process performed on the feature parameter sample set P after the dimension reduction in the step (4) includes the following steps:
1) for the original data set P after dimensionality reduction, determining the value of coefR, calculating a Density adjustment parameter sigma and a Density reachable distance R, and calculating the Density value sensitivity (P) of each data object in Pi) To obtain the sensitivity (P) in Pi) The largest partDensity Attractor:
Figure GDA0002950232510000141
Figure GDA0002950232510000142
Figure GDA0002950232510000143
in the formula (I), the compound is shown in the specification,
Figure GDA0002950232510000144
representing point PiTo point PjThe Euclidean distance of (a) is,
Figure GDA0002950232510000145
coefR (0), an average of the distances between points in the sample<coefR<1) The original adjustment coefficient of the density reachable distance is obtained, and n is the total number of samples;
2) scanning data objects in an original data set P, and allocating a density Attractor and data objects with reachable density to a first cluster C0And deleting the cluster object from the original data set;
3) for the remaining dataset, look up another density AttractoriCalculating the adaptive density reachable distance Radap,iAssign the density attractor and the data object whose density is reachable to another cluster CiDeleting the cluster object from the original data set, repeating the above steps, and finally putting the cluster with less data objects into an abnormal value or noise group, and obtaining an original cluster result Ci(i ═ 0,1, …, k) ∈ P, outlier O ∈ P:
Radap,i＝αR
where α is the adjustment factor, the formula is as follows:
Figure GDA0002950232510000146
in an embodiment of the present invention, the incremental clustering method in step (5) is:
1) for the preprocessed incremental data set delta P, combining the delta P with the original outlier O, and searching each cluster C in the O U delta PiUpdating clustering result C for data objects with reachable densityi(i ═ 0,1, …, k) ∈ (P ≧ Δ P) when the remaining dataset is
Figure GDA0002950232510000151
2) For the remaining data sets
Figure GDA0002950232510000152
The data object is subjected to clustering analysis to obtain the data object possibly existing in
Figure GDA0002950232510000153
Cluster Δ C in (1)i(i ═ k +1, …, k ', k ' ≧ k), and the update clustering result is C 'i(i＝0,1,…,k′)＝Ci(i＝0,1,…,k)∪ΔCiAnd an outlier O';
3) if it is
Figure GDA0002950232510000154
Detecting whether the O 'contains a data object in the incremental sample delta P, if the O' contains a new product parameter sample, comparing the distance from the new sample point to each point in each cluster, and marking the new sample point to the cluster where the closest point is located in order to realize energy consumption prediction of the new product;
4) after other incremental data sets are preprocessed, deleting the clustering mark in the existing cluster O', and repeatedly executing the steps 1) to 3);
5) outputting a clustering result C'i(i ═ 0,1, …, k ') and the outlier O', and the category to which the new product belongs.
In the embodiment of the present invention, the method for constructing the single product energy consumption prediction model in step (9) is as follows:
1) the dimension of the ADE individual is equal to the sum of the weight of the LSTM neural network and the threshold number, the initial evolution time G is 0, the population size N, the cross probability f and the variation probability cr, and the root mean square error RMSE is adopted as a fitness function:
Figure GDA0002950232510000155
Figure GDA0002950232510000156
Figure GDA0002950232510000161
in the formula, GenM is the maximum iteration number, G is the current iteration number, and alpha and beta are [0.5,1]Constant within the range, ytK is the true value, and k is the number of data objects;
2) calculating the fitness value (RMSE) of each individual in the population;
3) if the minimum RMSE in the current population meets the requirement or the current iteration time G is GenM, the ADE terminates the iteration, obtains the best individual and executes the step 5), otherwise, the step 4) is continuously executed;
4) obtaining a new group according to the self-adaptive intersection, the self-adaptive variation and the selection operation, setting G as G +1, and returning to execute the step 3);
5) obtaining optimal individuals as initial connection weights and thresholds of the LSTM based on ADE optimization; and training the LSTM by using the training samples to further obtain the optimal network.
In the embodiment of the present invention, the method for updating the LSTM prediction model by adding the new sample in step (11) includes:
1) for each newly-increased energy consumption sample, firstly, judging which cluster c belongs toiThe ith cluster corresponds to the ith LSTM model; after the model is determined, the model parameters are updated on the basis of historical data training of the model, namely, the parameters after the historical data training are used for initializing the network, and the strong explanatory variables of the new sample are input into the prediction modelIn the method, the predicted value of the new sample is obtained through the forward calculation of the LSTM
Figure GDA0002950232510000162
Will predict the value
Figure GDA0002950232510000163
The error from the actual value y is added to the original error function J (θ);
Figure GDA0002950232510000164
wherein the parameter θ ═ Wf,Wi,Wc,Wo,bf,bi,bc,bo)
2) Updating the model parameter values according to the updated error function J (theta) back propagation:
θ′＝(Wf-λ*Δwf,…,Wo-λ*Δwo,bf-λ*Δbf,…,bo-λ*Δbo)
wherein λ is a learning rate, Δ wf、ΔwoAnd Δ bf、ΔboGradient matrices and vectors of weights and offsets of the neurons, respectively.
Description of data
The experimental data are derived from data of a hub manufacturing company, which has intelligent instruments installed in the production line to obtain the electric energy, water and natural gas consumed in the production process. In the experiment, for example, single-product power consumption is predicted, as the available data amount is small, part of data is simulated by Python, relevant data of 725 wheel types in 2015 year 1 month to 2018 year 12 month are obtained at present, 475 wheel types are randomly selected to be used as historical product samples in the experiment, and the rest 250 wheel types are regarded as prediction results of newly-added products for verification models.
The characteristic parameter samples of the selected 475 wheel types are used as an original data set P, and the residual parameter samples of the 250 wheel types are divided into five groups in an equal amount at random for being used as an incremental data set deltaP1、ΔP2、ΔP3、ΔP4、ΔP5Incremental clustering was performed and the specific characteristic parameter sample description is shown in table 1 below.
TABLE 1 characterisation of the parameters data
Figure GDA0002950232510000171
And on the basis of the clustering result, constructing an LSTM increment updating hub unit consumption prediction model based on the ADE. The 725 wheel-type order data, production parameters, key consumable part parameters and energy consumption data form an energy consumption sample, wherein the energy consumption data is the single-product power consumption, 34,800 samples are counted, and 475 historical products comprise 22,800 samples. And randomly dividing samples contained in each cluster of the historical product into training samples and testing samples according to the ratio of 3:1, and constructing an energy consumption prediction model of each cluster. In the training process, the training set of each cluster is equally divided into four groups, wherein one group is used for constructing a prediction model, and the remaining three groups are added to the existing model for incremental learning. Specific energy consumption data and its associated influencing factors are shown in table 2 below.
TABLE 2 energy consumption samples
Figure GDA0002950232510000181
1. Model parameters and Structure
(1) Cluster analysis
The characteristic spoke front modeling (p9), the center disc modeling (p10), the hub material (p11) and the manufacturing process (p12) are all character type, and the standardization needs to be changed into numerical data.
The method comprises the steps of dividing a sample into historical product samples and newly added product samples in a random sampling mode, wherein 475 historical products are obtained, the rest newly added products are added into a clustering model in five times for incremental updating, the coefR value is 0.6, 20 times of random sampling are carried out in the experiment, the hubs can be finally divided into 5 types, and the clustering result is shown in fig. 4.
One of the experimental results was selected for illustration: carrying out cluster analysis on 475 randomly-extracted wheel-type dimensionality-reduced parameter samples, and finally classifying the samples into 6 classes C-1、C0、C1、C2、C3、C4There are 13, 261, 53, 27, 32, 89 sample points, respectively, where C-1Is a cluster containing noise points, C-1As the ion cluster O. The rest 250 hub parameter samples are subjected to incremental updating on the clustering result for 5 times, and the incremental updating samples can be searched for clustering C0、C1、C2、C3、C4Density reachable data objects, Final Cluster C0、C1、C4Respectively 303, 78, 185 samples, and an off-cluster O of 9 sample points, C2、C3No new sample was added.
(2) Analysis of key influence factors of energy consumption
1) Correlation analysis
And (3) preliminarily judging whether the single-product power consumption and each influence factor have linear correlation by using the Pearson coefficient, wherein the calculation result is shown in a table 3.
TABLE 3 Person coefficient 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
Therefore, X1, X4 and X5 are almost irrelevant to the single-product energy consumption, X7 is strongly and negatively correlated to the single-product energy consumption, and the rest variables are positively correlated. Therefore, by this result, three weak correlation factors of X1, X4, and X5 can be removed.
2) Adaptive-Lasso variable selection
Then, the Adaptive-Lasso algorithm is adopted for the remaining variables to carry out the second variable screening, and the results are shown in table 4:
TABLE 4 Adaptive-Lasso coefficient 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 can be known that if the coefficients of X6, X7, X11, and X10 are 0, these variables are eliminated during model building, and the remaining variables X2, X3, X8, and X9 are used as key influencing factors of power consumption of single products.
3) Prediction of strongly explanatory factors for new products
Through the analysis, X2, X3, X7 and X9 are strong explanatory factors of energy consumption, X2 is order quantity, X3 is the number of different product types produced by the same production line in the same production period, so that X2 and X3 of newly added products can be set manually according to actual conditions, X8 and X9 need to be obtained by BP neural network prediction, 12 characteristic parameters of the hub are used as input variables, and X8 and X9 are outputs.
In the experimental process, models are trained in a K-Fold cross validation mode, original data are divided into K groups (K-Fold), each subset data is subjected to a primary validation set, and the rest K-1 groups of subset data are used as training sets to obtain K models. And evaluating the results of the k models in a verification set respectively, and finally selecting the model with the minimum error. In the experiment, k is 4, namely, the data is divided into 4 parts, k-1 is 3 parts of the data to be trained, the rest parts are used for verification, the network is trained for 4 times in total, and each part of the data has an opportunity to be used as a verification set. In the experimental process, the number of cycles of each training is gradually increased to 300, namely the epoch is 300, and the error of the verification data after each training is recorded. During each training, the learning rate is dynamically adjusted, and when learning stagnates, the learning rate is reduced by 0.1 times.
For the determination of the number of hidden layers, by setting BP with the same other parameters, the number of hidden layers is gradually increased from 1 layer to 5 layers of RMSE of the predicted value, as shown in table 5 below, when a hidden layer is a 1 layer, both RMSE values of X8 and X9 are smaller, and when the number of hidden layers is gradually increased from 2 to 5, the RMSE value fluctuates within a smaller range, so 1 hidden layer may be selected.
TABLE 5 Effect of hidden layer number on the prediction accuracy of Key influencing factor
Figure GDA0002950232510000201
Setting BP other parameters to be same for determining the number of hidden layer nodes, and setting BP other parameters to be same according to a formula
Figure GDA0002950232510000202
Searching, wherein x is the number of nodes of an input layer, namely 12, y is the number of nodes of an output layer, namely 2, a is a constant which is more than or equal to 1, and z is the number of nodes of a hidden layer, so that the initial value of the number of nodes of the hidden layer can be set to be 5 and gradually increased, errors of predicted values are recorded, experimental results are shown in figures 5 and 6, when the number of nodes of the hidden layer is increased from 5 to 16, RMSE of the light finishing efficiency and the rejection rate is gradually reduced, and when the number of nodes of the hidden layer is continuously increased, RMSE of the light finishing efficiency and the rejection rate is gradually increased,the number of hidden layer nodes is taken to be 16.
The experimental results show that the BP neural network for predicting key influence factors of newly-added products has a 3-layer structure, wherein the number of nodes of an input layer is the same as the dimension of the characteristic parameters of the hub, namely 12, the number of nodes of a hidden layer is 16, and the number of nodes of an output layer is 2. And (3) operating the model for 10 times, taking the average value of the predicted values of X8 and X9 as a final predicted value, and setting X2 and X3 variables of each newly added product for inputting a verification energy consumption prediction model.
(3) Energy consumption prediction model for each cluster
According to the result of the cluster analysis, the historical products are aggregated into 5 types, energy consumption samples contained in each cluster of the historical products are randomly divided into training samples and testing samples according to the ratio of 3:1, and an energy consumption prediction model of each cluster is constructed. In the training process, the training set of each cluster is equally divided into four groups, wherein one group is used for constructing a prediction model, and the remaining three groups are added to the existing model for incremental learning. The number of nodes of the input layer of each model is the same as the dimension of the strong explanatory factor, namely 4, and the number of nodes of the output layer is 1, namely, the single-product energy consumption value under the influence of each characteristic is output. In addition, the population size of the ADE search part is set to be 50, the iteration times are 100, and alpha and beta are respectively 0.8 and 0.5.
1) Determining LSTM layer number
According to cluster C0、C1、C2、C3、C4And constructing a single-product energy consumption prediction model of the product by using the contained single-product energy consumption data. Setting other parameters of the model to be the same, gradually increasing the LSTM layer of the neural network from 1 layer to 10 layers, recording the change condition of the RMSE of the model predicted value, and the experimental result is shown in FIG. 7. For class C0And C4The prediction model of (1) shows that the RMSE value gradually decreases when the number of LSTM layers gradually increases from 1 to 3, and the RMSE fluctuates in a small range when the number of LSTM layers continues to increase, so that C0And C4The predictive model of (a) can be determined as a 3-layer LSTM structure. And for C1、C2、C3The prediction model of (2) shows that the RMSE values are all significantly smaller when the LSTM is increased from 1 layer to 2 layers, and when the LSTM is increased from 1 layer to 2 layersAs the LSTM layer continues to increase, the RMSE fluctuates within a smaller range with no apparent tendency to decrease, so C1、C2、C3The LSTM layer of the predictive model of (1) is two-layered.
2) Determining number of LSTM layer nodes
For determining the number of LSTM layer nodes in each cluster, setting other parameters of each model to be the same, only changing the number of nodes, and obtaining each cluster C through experiments0、C1、C2、C3、C4The ratio of the nodes of each layer of the prediction model is as follows: the ratio of the first layer node to the second layer node is 4:2:1, 5:2 and 4:3:1, the change of the model prediction error is recorded, the number of the other layers of nodes is set by proportion, and the experimental result is shown in fig. 8. For cluster C0For the prediction model of (1), when the number of nodes is increased from 4 to 24, each error index is obviously reduced, when the number of nodes is increased from 28 to 36, the error index has a tendency of reducing but only fluctuating in a small range, and when the number of nodes is continuously increased, the error index has a tendency of rising, so that the first layer adopts 24 nodes. In the same way, C1、C2、C3、C4The first levels of the predictive model of (1) are 16, 20, respectively.
2. Analysis of results
Cluster C0、C1、C4And adding new products, so that a comparison test adopts a prediction model corresponding to the three clusters for analyzing the prediction precision of the energy consumption of the single products of the new products. Cluster C0、C1、C4The energy consumption samples of the contained historical products are randomly divided into training samples and testing samples according to the proportion of 3: 1. In the training process, the training set is equally divided into four groups, one group is used for building a prediction model, and the remaining three groups are added to the existing model for incremental learning. The energy consumption prediction method based on the combination of the incremental density clustering and the LSTM provided by the invention is used for carrying out incremental learning comparison with the DE-LSTM method based on LSTM, BP, SVR and non-incremental learning, wherein the input and the output of the traditional algorithm adopt the same input and output of the method, the prediction precision and the running time of 10 experiments are recorded, the average value is obtained, and the comparison result is shown in the following table 6.
TABLE 6 model prediction accuracy comparison
Figure GDA0002950232510000231
As can be seen from the table, the LSTM hub single-product energy consumption prediction model based on incremental clustering provided by the invention is basically superior to other algorithms in prediction precision and training time. From the prediction accuracy, the RMSE of the method in the training stage, the testing stage and the newly-added product energy consumption prediction value is minimum, the RMSE of the training set is as low as 0.016524, the RMSE of the model training stage is reduced by 0.012545, 0.015129 and 0.007131 in average compared with LSTM, BP and DE-LSTM, and is reduced by 0.019806 in average compared with SVR, so that the LSTM hub single-product energy consumption prediction model based on incremental clustering is obviously improved due to the fact that ADE is adopted to optimize model initial parameters, more importantly, an incremental learning strategy is adopted, and the time-sequence characteristic of energy consumption characteristics is considered. From the runtime perspective, the method has more runtime than SVR, because the method needs to build a multi-layer LSTM neural network, and has less training time and runtime relative to other algorithms, because other algorithms cannot perform incremental learning, and need to retrain each time a training set is added, thus increasing runtime.
Next, analysis is performed on the training situation of each time of adding training data, as shown in fig. 9-11, when the historical energy consumption samples contained in each cluster are sequentially added to the model in four times, the training precision change diagram at each time of adding is also compared with the LSTM, BP, SVR, and DE-LSTM algorithm without incremental learning, and as can be seen from fig. 9-11, the RMSE value of the algorithm at the first training time is not obviously different, and is obviously reduced with the addition of incremental data relative to the RMSE values of other algorithms without incremental learning.
Fourth, conclusion
In order to solve the problem that a traditional machine learning model cannot realize unit consumption prediction and incremental learning of a new product, the invention provides an LSTM hub unit energy consumption prediction method based on incremental clustering, dynamic incremental density clustering based on PCA is adopted to analyze characteristic parameters of a hub, so that a historical product category similar to the newly added product is obtained, meanwhile, strong explanatory factors of unit energy consumption are screened out through a Person coefficient and an Adaptive-Lasso algorithm, the strong explanatory factors of the newly added product are predicted by combining a BP algorithm, an energy consumption prediction model is separately constructed for each product according to a clustering result, an Adaptive differential evolution algorithm is adopted to optimize LSTM, an incremental learning strategy is introduced to consider the time sequence characteristics of a sample, when the energy consumption sample of the newly added product is obtained, the model can be subjected to incremental updating, and time cost and space consumption are saved. Through experimental analysis, the effectiveness of the proposed LSTM hub single-product energy consumption prediction method based on incremental clustering is verified, the RMSE of energy consumption prediction is reduced to 0.016524, compared with other algorithms without incremental learning, the RMSE is averagely reduced by 0.013653, and the reliable prediction of single-product energy consumption is realized. The main advantages are as follows:
(1) aiming at the problem that the product classification is difficult due to various complex and variable hub characteristic parameters, the dynamic incremental density clustering based on the PCA is provided, the method extracts the main characteristic components of the parameters through the PCA and keeps the uniqueness of the product, and then the clustering algorithm is utilized to obtain the historical single product category similar to a new product;
(2) aiming at the problem of importance analysis of energy consumption influence factors in the hub production process, an energy consumption influence factor system is constructed by integrating order data, production data and key consumable part parameters, strong interpretability factors of single-product energy consumption are extracted through Pearson coefficients and Adaptive-Lasso algorithm, the dimension of input data is reduced, and a BP neural network is used for predicting the strong interpretability factor value of a new product;
(2) aiming at the problems that the traditional machine learning model cannot predict new product energy consumption and cannot perform incremental learning, the method provides an LSTM incremental updating hub unit consumption prediction model based on ADE, the model optimizes model parameters by using an ADE algorithm, sums errors between predicted values and actual values of new samples and overall errors of a historical model, and iteratively updates the model parameters according to an error minimization method to ensure the incremental learning capability of the model, so that the model can be suitable for actual conditions with continuously increased data in production, and space occupation and time cost are saved.

Claims (5)

1. An LSTM hub single-product energy consumption prediction method based on incremental clustering is characterized by comprising the following steps:
(1) collecting characteristic parameter samples of hubs of different models, wherein the characteristic parameter samples comprise 12 characteristic parameters: the wheel rim comprises a wheel rim diameter, a wheel rim width, a center hole distance, a bolt hole number, a pitch circle diameter, an offset distance, weight, a wheel spoke number, a wheel spoke front face model, a center disc surface model, a wheel hub material and a manufacturing process;
(2) collecting energy consumption samples consisting of order data, production data, key consumable part parameters and unit product energy consumption data of hubs of different models; the order data, the production data and the key consumable part parameters form an energy consumption influence factor system, the order data comprises a single product model number, a single product order amount and a total number of product types, the production data comprises production time, raw material input amount, equipment running time, production efficiency and rejection rate, the production efficiency is considered by two aspects of machining efficiency and finishing efficiency, and the machining efficiency is unit: piece/hour; polishing efficiency unit: minutes per piece; the key consumable parameters comprise the usage amount of a drill bit and the usage amount of a cutter;
(3) labeling character type parameters of the hub characteristic parameter sample in the step (1), and reducing the characteristic parameter sample to two dimensions by adopting principal component analysis to obtain a data set P;
(4) then, clustering operation is carried out on the P by using a dynamic incremental density clustering algorithm to obtain an original clustering result CiE.g., P, i ═ 0,1,. k, and the outlier O e.g., P;
(5) when a new characteristic parameter sample exists, processing is carried out by adopting an incremental clustering method, an incremental data set delta P is obtained after the new characteristic parameter sample is preprocessed in the step (3), a data object which can reach the original clustering density in the O U delta P is searched, a clustering result is updated, and a cluster C 'is output'i＝Ci∪ΔCiWherein C isiWherein i ═ 0, 1.,. k, C'iWhere i is 0,1iK ', k ' is not less than k, and is separated from cluster O ' and newly added productThe category of the genus;
(6) analyzing strong explanatory factors of the energy consumption of the single product according to the energy consumption influence factor system in the step (2), eliminating weak relevant and irrelevant factors in the energy consumption influence factor system by using a Pearson coefficient, and then performing variable selection by using an Adaptive-Lasso algorithm to obtain strong explanatory variables of the energy consumption of the single product of the hub;
(7) standardizing the hub characteristic parameters subjected to labeling processing in the step (3), the strong interpretable variable obtained in the step (6) and the energy consumption data of the unit product in the step (2);
(8) taking the hub characteristic parameters subjected to the standardization processing in the step (7) as input, taking the strong interpretable variable subjected to the standardization processing in the step (7) as output, and constructing a BP (back propagation) prediction model of the new hub strong interpretable variable to predict the strong interpretable variable value of the new hub;
(9) constructing a single-product energy consumption prediction model of each cluster according to the clustering result in the step (4), randomly dividing each cluster energy consumption sample into a training set and a testing set according to the ratio of 3:1, dividing the training set into four groups, performing the step (10) on one group, and sequentially taking the rest groups as a sample increment set;
(10) constructing a clustered LSTM single-product energy consumption prediction model, taking the strong interpretability variable of the training set subjected to the standardization in the step (7) as input, taking the single-product energy consumption subjected to the standardization in the step (7) as output, and realizing the optimization of LSTM parameters by using an ADE algorithm in the model construction;
(11) in each cluster LSTM single-product energy consumption prediction model, carrying out incremental updating on the model by using the sample increment set in the step (9);
(12) and (3) testing the prediction model of each cluster updated in the step (11) according to the test set of each cluster in the step (9), evaluating the model, processing the characteristic parameter samples of the product to be predicted in the steps (3) to (5) and obtaining the strong explanatory variable of the new hub through the BP prediction model in the steps (6) to (8), predicting by using the LSTM single-product energy consumption prediction model of each cluster updated in the step (11), and outputting the predicted value of the energy consumption of the newly-increased product.
2. The LSTM hub singles energy consumption prediction method based on incremental clustering of claim 1, wherein the clustering process of the feature parameter sample set P after dimension reduction in the step (4) comprises the following steps:
1) for the original data set P after dimensionality reduction, determining the value of coefR, calculating a Density adjustment parameter sigma and a Density reachable distance R, and calculating the Density value sensitivity (P) of each data object in Pi) Obtaining the Density value Density (P) in Pi) Maximum data object local density Attractor:
Figure FDA0002979836660000031
Figure FDA0002979836660000032
Figure FDA0002979836660000033
in the formula (I), the compound is shown in the specification,
Figure FDA0002979836660000034
representing point PiTo point PjThe Euclidean distance of (a) is,
Figure FDA0002979836660000036
the coefR is the average value of the distances between each point in the sample, 0-1 is the original adjustment coefficient of the density reachable distance, and n is the total number of the samples;
2) scanning data objects in the original data set P, and attaching a density Attractor0And data objects whose density is reachable are assigned to the first cluster C0And deleting the cluster object from the original data set;
3) for the remaining dataset, look up another density AttractoriCalculating the adaptive density reachable distance Radap，iAssign the density attractor and the data object whose density is reachable to another cluster CiDeleting the cluster object from the original data set, repeating the above steps, and finally putting the cluster with less data objects into an abnormal value or noise group, and obtaining an original cluster result CiBelongs to P, and the outlier O belongs to P:
Radap，i＝αR
where α is the adjustment factor, the formula is as follows:
Figure FDA0002979836660000035
3. the LSTM hub singles energy consumption prediction method based on incremental clustering of claim 1, wherein the incremental clustering method in step (5) is:
1) for the preprocessed incremental data set delta P, combining the delta P with the original outlier O, and searching each cluster C in the O U delta PiUpdating clustering result C for data objects with reachable densityiE (P ≧ Δ P), when the remaining dataset is
Figure FDA0002979836660000041
2) For the remaining data sets
Figure FDA0002979836660000042
The data object is subjected to clustering analysis to obtain the data object possibly existing in
Figure FDA0002979836660000043
Cluster Δ C in (1)iUpdating the clustering result to be C'i＝Ci∪ΔCiAnd an outlier O';
3) if it is
Figure FDA0002979836660000044
Detecting whether the O' contains the data object in the incremental data set delta P, if soIf the new product parameter samples are contained, comparing the distances from the data object to each point in each cluster, and marking the data object to the cluster where the closest point is located in order to realize energy consumption prediction of the new product;
4) after other incremental data sets are preprocessed, deleting the clustering mark in the existing cluster O', and repeatedly executing the steps 1) to 3);
5) outputting a clustering result C'iAnd an outlier O', and the category to which the new product belongs.
4. The LSTM hub single product energy consumption prediction method based on incremental clustering of claim 1, wherein the single product energy consumption prediction model in the step (10) is constructed by the following steps:
1) the dimension of the ADE individual is equal to the sum of the weight of the LSTM neural network and the threshold number, the initial evolution time G is 0, the population size N, the cross probability f and the variation probability cr, and the root mean square error RMSE is adopted as a fitness function:
Figure FDA0002979836660000045
Figure FDA0002979836660000046
Figure FDA0002979836660000051
in the formula, GenM is the maximum iteration number, G is the current iteration number, and alpha and beta are [0.5,1]Constant within the range, ytK is the true value, and k is the number of data objects;
2) calculating the fitness value (RMSE) of each individual in the population;
3) if the minimum RMSE in the current population meets the requirement or the current iteration time G is GenM, the ADE terminates the iteration, obtains the best individual and executes the step 5), otherwise, the step 4) is continuously executed;
4) obtaining a new group according to the self-adaptive intersection, the self-adaptive variation and the selection operation, setting G as G +1, and returning to execute the step 3);
5) obtaining optimal individuals as initial connection weights and thresholds of the LSTM based on ADE optimization; and training the LSTM by using the training samples to further obtain the optimal network.
5. The LSTM hub singles energy consumption prediction method based on incremental clustering of claim 1, wherein the updating method of the LSTM prediction model by the newly added samples in step (11) comprises:
1) for each newly-increased energy consumption sample, firstly, judging which cluster c belongs toiThe ith cluster corresponds to the ith LSTM model; after the model is determined, the model parameters are updated on the basis of the historical data training of the model, namely, the parameters after the historical data training are used for initializing the network, the strong explanatory variable of the new sample is input into the prediction model, and the predicted value of the new sample is obtained through the forward calculation of the LSTM
Figure FDA0002979836660000052
Will predict the value
Figure FDA0002979836660000053
The error from the actual value y is added to the original error function J (θ);
Figure FDA0002979836660000054
wherein the parameter θ ═ Wf，Wi，Wc，Wo，bf，bi，bc，bo)
2) Updating the model parameter values according to the updated error function J (theta) back propagation:
θ′＝(Wf-λ*Δwf，...，Wo-λ*Δwo，bf-λ*Δbf，...，bo-λ*Δbo)
wherein λ is a learning rate, Δ wf、ΔwoAnd Δ bf、ΔboGradient matrices and vectors of weights and offsets of the neurons, respectively.
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 CN110070145A (en) 2019-07-30
CN110070145B true 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)

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
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
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
CN111832839B (en) * 2020-07-24 2021-04-30 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
CN112116140B (en) * 2020-09-10 2022-09-20 同济大学 Building energy consumption prediction method based on twin model
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
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
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
CN114202065B (en) * 2022-02-17 2022-06-24 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
CN115167279B (en) * 2022-09-07 2022-11-29 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
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
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
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

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
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

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10248332B2 (en) * 2016-10-07 2019-04-02 Prophetstor Data Services, Inc. Method for extending life expectancy of disks in cloud-based service system and system using the same
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
CN109063416B (en) * 2018-07-23 2019-08-27 太原理工大学 Gene expression prediction technique based on LSTM Recognition with Recurrent Neural Network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
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
Energy Load Forecasting using Deep Learning Approach-LSTM and GRU in Spark Cluster;Sumit Kumar等;《2018 Fifth International Conference on Emerging Applications of Information Technology (EAIT)》;20180924;第1-4页 *
Short- Term Prediction of Wind Power Based on Adaptive LSTM;Gang Xu等;《2018 2nd IEEE Conference on Energy Internet and Energy System Integration (EI2)》;20181220;第1-5页 *
一种基于LSTM神经网络的短期用电负荷预测方法;陈佳等;《电力信息与通信技术》;20170930;第19-25页 *
基于 GAN 技术的自能源混合建模与参数辨识方法;孙秋野等;《自动化学报》;20180531;第901-914页 *
改进PSO-BP网络预测模型在造纸能耗预测中的应用;陈周林等;《轻工科技》;20190108;第91-94页 *

Also Published As

Publication number Publication date
CN110070145A (en) 2019-07-30

Similar Documents

Publication Publication Date Title
CN110070145B (en) LSTM hub single-product energy consumption prediction based on incremental clustering
CN109583635B (en) Short-term load prediction modeling method for operational reliability
CN113298230B (en) Prediction method based on unbalanced data set generated against network
CN109492748B (en) Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network
CN110571792A (en) Analysis and evaluation method and system for operation state of power grid regulation and control system
CN112364560B (en) Intelligent prediction method for working hours of mine rock drilling equipment
Mehmanpazir et al. Development of an evolutionary fuzzy expert system for estimating future behavior of stock price
CN112330052A (en) Distribution transformer load prediction method
CN111832839B (en) Energy consumption prediction method based on sufficient incremental learning
CN112288157A (en) Wind power plant power prediction method based on fuzzy clustering and deep reinforcement learning
CN114117852B (en) Regional heat load rolling prediction method based on finite difference working domain division
CN106599436A (en) User in-room behavior prediction method for office building
Berneti Design of fuzzy subtractive clustering model using particle swarm optimization for the permeability prediction of the reservoir
CN109214610A (en) A kind of saturation Methods of electric load forecasting based on shot and long term Memory Neural Networks
CN117076691A (en) Commodity resource knowledge graph algorithm model oriented to intelligent communities
Wu et al. A forecasting model based support vector machine and particle swarm optimization
Yang Combination forecast of economic chaos based on improved genetic algorithm
CN110059871B (en) Photovoltaic power generation power prediction method
CN116956160A (en) Data classification prediction method based on self-adaptive tree species algorithm
CN117096860A (en) Overhead transmission line current-carrying capacity interval prediction method and equipment based on LSSVM model
CN116303386A (en) Intelligent interpolation method and system for missing data based on relational graph
CN113656707A (en) Financing product recommendation method, system, storage medium and equipment
CN113537553A (en) Regional power distribution network load demand hybrid prediction method
Du et al. Short-term power load forecasting based on spark platform and improved parallel ridge regression algorithm
Adesiyan Performance prediction of production lines using machine learning algorithm

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