CN117709536B - Accurate prediction method and system for deep recursion random configuration network industrial process - Google Patents

Accurate prediction method and system for deep recursion random configuration network industrial process Download PDF

Info

Publication number
CN117709536B
CN117709536B CN202311743288.7A CN202311743288A CN117709536B CN 117709536 B CN117709536 B CN 117709536B CN 202311743288 A CN202311743288 A CN 202311743288A CN 117709536 B CN117709536 B CN 117709536B
Authority
CN
China
Prior art keywords
output
neurons
layer
pool
network
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
CN202311743288.7A
Other languages
Chinese (zh)
Other versions
CN117709536A (en
Inventor
王殿辉
党纲
Original Assignee
东北大学
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 东北大学 filed Critical 东北大学
Priority to CN202311743288.7A priority Critical patent/CN117709536B/en
Publication of CN117709536A publication Critical patent/CN117709536A/en
Application granted granted Critical
Publication of CN117709536B publication Critical patent/CN117709536B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • 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"
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • 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
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
    • G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
    • G16C20/10—Analysis or design of chemical reactions, syntheses or processes
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
    • G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
    • G16C20/70—Machine learning, data mining or chemometrics
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/003—Load forecast, e.g. methods or systems for forecasting future load demand

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Chemical & Material Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Crystallography & Structural Chemistry (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Development Economics (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Analytical Chemistry (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Power Engineering (AREA)
  • Neurology (AREA)
  • Game Theory and Decision Science (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)

Abstract

The invention belongs to the technical field of dynamic modeling, and particularly relates to a method and a system for accurately predicting a deep recursion random configuration network industrial process, which can utilize stored historical information to perform modeling, only input data at the current moment, and do not need to acquire all order delay data. Under the condition that the reduction of the root mean square error of the neural network model output is met, a proper neuron is found to serve as a reserve pool candidate neuron through the inequality constraint condition of a random configuration algorithm. And reassigning the connection weights of the newly added neurons, only reserving the network connection weights of the newly added neurons to the newly added neurons and the original neurons, and setting the network connection weights of the original neurons to the newly added neurons to be zero, namely, no connection. After determining the newly added neuron, calculating to obtain an output weight by using a least square method, and judging whether the construction of the reserve pool of the current layer is completed. And then configuring a next layer of reserve pool, and judging whether the network is constructed or not through the maximum layer number, the maximum allowable neuron number of the reserve pool and the maximum allowable output error.

Description

Accurate prediction method and system for deep recursion random configuration network industrial process
Technical Field
The invention belongs to the technical field of dynamic modeling, and particularly relates to a method and a system for accurately predicting a deep recursion random configuration network industrial process.
Background
Complex systems in industrial processes have a multi-variable dynamic evolution behavior, and each input variable has a certain delay, i.e. the effect of the input at the current moment on the system is reflected after a period of time. However, due to the influence of factors such as process environment, external disturbance and the like, it is difficult to capture all order delay information related to input variables, and input orders at different moments are continuously changed. It is important how to accurately model such order-uncertain nonlinear dynamic systems with effective use of known variable information.
Through the above analysis, the problems and defects existing in the prior art are as follows:
(1) The recurrent neural network and the long-short-term memory network can store a part of history information due to feedback connection among the neurons of the recurrent neural network and the long-short-term memory network, and can be used for solving the problem of nonlinear dynamic system modeling with uncertain orders. However, the two networks need to adjust the network connection weight and the bias according to an error gradient descent method in the training process, so that the training process has the problems of large calculation amount, long time consumption, easy local optimization, gradient disappearance or explosion and the like, and low implementation efficiency.
(2) Feedback connection exists among all neurons of the echo state network, and compared with a recurrent neural network, the echo state network calculates an output weight through a least square method, and the input weight and the reserve pool connection weight are randomly given and are not changed. However, the echo state network has the defects of blind selection of a network structure, sensitive weight parameters, incapability of guaranteeing global approximation property of a model and the like, and the range of weight and bias cannot be adaptively adjusted according to input data, so that the prediction accuracy of the model is affected.
(3) The reserve pool of the recursion random configuration network also adopts random connection, and inequality constraint is introduced, and the reserve pool neurons are added in a supervision mode, so that the problems can be effectively avoided. However, compared with the multi-layer network model, under the condition that the total neuron numbers are similar, the calculation complexity of the single-layer echo state network and the recursive random configuration network is high, and the model training speed is slow.
In the industrial field, the problems and drawbacks of the above-mentioned techniques lead to the following problems:
1. The real-time performance is not enough: because the recurrent neural network and the long and short-term memory network (LSTM) need to adjust the network connection weight and the bias according to an error gradient descent method in the training process, the training process has large calculation amount and long time consumption, and is not suitable for industrial application requiring quick response, such as real-time monitoring and fault detection.
2. The prediction precision is low: the echo state network and the recursion random configuration network have the problems that weight parameters are sensitive, global approximation property of a model cannot be guaranteed and the like, so that the prediction accuracy of the model is low, and the requirements cannot be met for industrial applications requiring high-accuracy prediction, such as quality control, energy system load prediction and the like.
3. The resource use efficiency is low: compared with a multi-layer neural network, the single-layer recursive random configuration network has higher calculation complexity and slow model training speed under the condition that the total neuron numbers are similar, so that the resource utilization efficiency is low.
4. Stability and robustness are insufficient: because the recurrent neural network and the long-short memory network are easy to be in local optimum, the problems of gradient disappearance or explosion and the like are generated, the stability and the robustness of the model are insufficient, and risks exist for industrial systems which need to stably run under various working conditions.
5. Model parameter adjustment is difficult: because the models can not adaptively adjust the weight and the bias range according to the input data, the models are difficult to adjust, and can not be quickly adapted to the changes of the industrial system, and the operation efficiency and the effect of the system are affected.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides a method and a system for accurately predicting the industrial process of a deep recursion random configuration network, which aim to solve the problem of modeling a nonlinear dynamic system with uncertain orders and improve the integral training speed and prediction accuracy of a model. The deep recursion random configuration network can utilize the stored historical information to perform modeling, only the data at the current moment is input, and all order delay data are not required to be acquired. Under the condition that the reduction of the root mean square error of the neural network model output is met, a proper neuron is found to serve as a reserve pool candidate neuron through the inequality constraint condition of a random configuration algorithm. Different from the random connection of the original reserve pool, the deep recursion random configuration network redistributes the connection weight of the newly added neuron, only retains the network connection weight of the newly added neuron to the original neuron and the network connection weight of the original neuron, and sets the network connection weight of the original neuron to the newly added neuron to zero, namely no connection. After determining the newly added neurons, calculating to obtain an output weight by using a least square method, and judging whether the reservoir of the current layer is constructed or not according to the maximum allowable neuron number and the maximum allowable output error of the reservoir. Then, the next layer of pool is configured in the same step, the pool is continuously pushed to the deeper layer of pool, and whether the network is built is judged through the maximum layer number, the maximum allowable neuron number of the pool and the maximum allowable output error.
The invention is realized in such a way that the accurate prediction method of the deep recursion random configuration network industrial process is adopted, and the technical scheme adopts the accurate prediction method based on the deep recursion random configuration network (DeepRSCN), and the dynamic system modeling with high adaptability and high prediction precision is realized by data preprocessing, network hierarchy initialization, random configuration weight and bias, selection and optimization of the reserve pool neurons in combination with inequality constraint conditions and the projection algorithm updated by real-time data. The scheme reduces the dependence on a large amount of tag data, remarkably improves the calculation efficiency, has good expansibility and instantaneity, and is suitable for dynamic modeling of a complex system requiring quick response and high-precision prediction.
Further, the method comprises the steps of:
S1, preprocessing data;
S2, initializing a deep recursion random configuration network;
S3, randomly configuring weights and biases, and constructing candidate neurons of the reserve pool by adopting a brand new method;
s4, selecting candidate neurons in a supervision mode according to inequality constraint conditions;
S5, determining optimal candidate neurons, and adding the optimal candidate neurons into a reserve pool;
s6, calculating the current layer network output weight by combining the target output and the reserve pool output;
s7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3-S6 until the error requirement is met or the maximum allowable number of neurons of the reserve pool of the current layer is reached;
S8, when the error requirement is not met, entering the next layer, and sequentially configuring the storage pool neurons of each layer according to S3-S7 until the error requirement is met or the maximum layer number requirement is met;
And S9, based on a projection algorithm, carrying out online updating on the trained output weight according to the real-time data, and calculating real-time output.
Further, S1 specifically includes: given a set of timing data, input samples: { u (1), u (2),..u (n max)}＝{(y(1),u(1)),(y(2),u(2)),...,(y(nmax),u(nmax)) }, target output: t= { T (2), T (3), T (n max +1) }, where y, u are the system output and the controlled input, respectively, the present invention predicts the next moment output y (n+1) only by the current moment input (y (n), u (n)), n max is the number of samples. Setting the maximum layer number as S, and the maximum allowable number of neurons in each layer of reserve poolThe maximum expected output error tolerance epsilon, the maximum candidate neuron generation number G max, and the input weight distribution parameter gamma= { lambda min,λmin+Δλ,...,λmax }.
Further, S2 specifically includes: initializing an output error vector e 0 =t, and outputting an error scaling factor 0 < r < 1 by a model, wherein the network model of different layers is as follows:
…
Where j=1, 2,.., And/>Is the input weight of the j layer, and the reserve pool is connected with the weight and the output weight matrix,/>Is the bias matrix, X (j) (n) is the pool state vector at time n, g is the activation function, X (j)＝[x(j)(1),x(j)(2),…x(j) (n) ] is the pool state matrix, and Y (j) is the network output of layer j.
Further, S3 specifically includes: randomly selecting an input weight from gamma= { lambda min,λmin+Δλ,...,λmax }, andPool connection weights/>And bias/>Substitution of activation functions(First layer) or(Layer j) obtaining G max group candidate neurons
Further, S4 specifically includes: substituting G max sets of candidate neurons into the inequality constraint of the random configuration algorithm:
Wherein the method comprises the steps of Is the output error of all pool neurons of each layer, when the single-layer structure is adopted: n sum = N, layer j: /(I)M is the output dimension, non-negative real sequence/>Satisfy/>And/>Candidate neurons satisfying the inequality constraint are screened out.
Further, S5 specifically includes: defining a set of variables: Candidate neurons satisfying the inequality constraint and maximizing ζ N+1 are selected as optimal pool candidate neurons.
Further, S6 specifically includes: adding the optimal reserve pool candidate neurons into a network model, and directly calculating the output weight of the neural network model according to the target outputCalculating the output root mean square error of the neural network model, taking the root mean square error as a loss function, and updating the output error/>, of the modelAnd pool neuron number n=n+1.
Further, S7 specifically includes: if the output root mean square error e 0||2 of the current network model is greater than the maximum expected output error tolerance epsilon and the pool neuron number N is less than the pool neuron maximum toleranceRepeating S3-S6; if the output root mean square error of the current network model, e 0||2, is greater than the maximum expected output error tolerance epsilon and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>Then the next layer is entered for continuous training, the number of the neurons in the reserve pool is updated to be N=0, the number of the neurons in the reserve pool is updated to be Chi Cengshu j=j+1, and the steps S3 to S6 are repeated; if the root mean square error of the current network model output i e 0||2 is less than the maximum expected output error tolerance epsilon, or j=s and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>And after the training is finished, obtaining a depth network model meeting the constraint condition of the random configuration theory.
Further, S9 specifically includes: based on a projection algorithm, online updating is carried out on the trained output weight according to real-time data, and real-time output is calculated:
where 0 < a < 1, c > 0 is two constants, g (n) is the reservoir output at time n, and e (n) is the output error before correction.
It is another object of the present invention to provide a dynamic modeling system for a complex system in an industrial process applying the deep recursive random configuration network industrial process accurate prediction method, comprising:
The preprocessing module is used for preprocessing data;
the initialization module is used for initializing the deep recursion random configuration network;
Constructing a neuron module, which is used for randomly configuring weights and biases and constructing candidate neurons of the reserve pool by adopting a brand new method;
The candidate neuron selection module is used for supervised selection of candidate neurons according to inequality constraint conditions;
an optimal neuron determination module that determines optimal candidate neurons to be added to the reservoir;
The output weight calculation module is used for calculating the current layer network output weight by combining the target output and the reserve pool output;
and the online updating module is used for online updating the trained output weight according to the real-time data based on the projection algorithm and calculating real-time output.
In combination with the technical scheme and the technical problems to be solved, the technical scheme to be protected has the following advantages and positive effects:
First, the present invention provides a method and system for accurate prediction of a deep recursive stochastic configuration network industrial process, using historical information stored in a reservoir to model when the input order is unknown. Compared with a recurrent neural network and a long-short-time memory network, the input weight and bias of the reserve pool neurons are randomly configured by introducing a supervision mechanism in the incremental construction process, and the iterative updating by counter propagation is not needed, so that the problems of easy local optimum, gradient disappearance or explosion and the like caused by gradient descent updating network parameters are avoided, and the network structure is determined in the configuration process; compared with a depth echo state network, the depth recursion random configuration network can adaptively adjust weight and bias according to input data, can theoretically ensure global approximation property of a model, and realizes zero error approximation of training data; compared with a single-layer recursion random configuration network, when the total node numbers are similar, the depth version can effectively reduce the training speed and the prediction accuracy of the model.
Secondly, the invention provides a precise prediction method and a precise prediction system for the deep recursion random configuration network industrial process, which can form more proper characteristic representation layer by layer, and have faster operation speed and higher prediction precision when the summary points are similar compared with a single-layer structure. The model framework designed by the invention is combined, so that the overall calculation amount of the algorithm can be reduced while the prediction accuracy is ensured, the method is suitable for application scenes with high real-time requirements, and has good application prospects in the fields of industrial artificial intelligence, intelligent medical treatment, intelligent traffic, unmanned driving and the like.
For the dynamic data modeling problem, people mostly stay in proving the existence of the global approximation capability of the model, but not from the aspect of constructivity. The invention provides a dynamic modeling algorithm and a system for a complex system in an industrial process, which are used for incrementally constructing a dynamic model with global approximation capability. The model does not need order identification, gradient descent is not needed to update the weight and the bias, the range of the weight and the bias can be adaptively adjusted, and the model is suitable for the problem of dynamic modeling with uncertain orders.
Third, the deep recursive random configuration network industrial process accurate prediction method has the following industrially significant technological advances:
1. enhancing model learning ability
Due to the adoption of the deep network structure, the model can learn more complex data distribution and dynamic characteristics, so that the prediction accuracy is improved. This is particularly important in complex industrial processes.
2. Dynamic and adaptive characteristics
The model can be self-adjusted and optimized in the running process through cyclic feedback and a reserve pool and online weight updating so as to adapt to dynamic changes in the industrial process.
3. Error monitoring and optimization
Through the multilayer reservoir and error feedback, the model can evaluate its performance in real time and adjust itself as needed to meet accuracy requirements.
4. Saving computing resources
The method is a random learning algorithm, the input weight and the bias are randomly given, gradient descent is not needed to update the weight, and compared with the traditional deep learning method, less calculation resources are needed, so that the method is very valuable in industrial environments with limited resources or quick response.
5. Enhancing approximation capability of a model
Candidate neurons are selected in a supervision mode through inequality constraint conditions, and the model has a universal approximation characteristic and can continuously approximate target output.
6. Hierarchical modeling
Because the model is multi-layered, the model can be subjected to layering modeling, not only can more complex system dynamics characteristics be captured, but also different optimization and control can be performed on different levels, and the flexibility of the model is improved.
7. Real-time performance
By updating the output weight on line, the model can adapt to new unknown data and conditions in real time, and is very important for the industrial field requiring real-time decision and control.
8. Reducing maintenance costs
The adaptive and real-time update feature reduces the complexity and cost of model maintenance because the model can self-adjust to dynamically changing industrial environments and conditions.
The accurate prediction method for the deep recursion random configuration network industrial process has remarkable industrial application potential in the aspects of improving model precision, saving computing resources, increasing self-adaptive capacity, realizing real-time control and the like.
Fourth, the significant technical progress brought by the accurate prediction method for the deep recursive random configuration network industrial process in the two embodiments provided by the present invention mainly includes:
1) Enhancement of model adaptation:
The storage pool is constructed through a supervision mechanism, the weight and the bias range are determined according to input data, the weight is output in real time by utilizing a projection algorithm, the model can dynamically adapt to system changes, and the modeling accuracy of an unknown or variable environment is improved.
2) And (3) improving prediction precision:
The weight and bias selection method driven by data is not blind random selection, and the deep learning structure can better capture nonlinear characteristics in data, so that the prediction accuracy is remarkably improved compared with the traditional linear model or shallow network.
3) Improvement of calculation efficiency:
the weight and bias are randomly configured in the initialization and iteration processes of the model, so that the weight calculation amount is reduced, and the calculation efficiency is improved.
4) Extensibility and flexibility:
The layering and modular design of the network structure allows the model to easily expand or modify the number of network layers and neurons according to different tasks.
5) Improvement of real-time performance:
The model can receive new data in real time and quickly adjust model parameters, so that dynamic modeling is more in line with actual application requirements, and especially in a scene with high real-time requirements.
In embodiments of chemical reaction process prediction, these technological advances can achieve more stable production processes and higher product quality. In the embodiment of power system load prediction, the scheduling and distribution of power resources can be optimized, the energy waste is reduced, and the running efficiency of a power grid is improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a method for accurate prediction of a deep recursive random configuration network industrial process provided by an embodiment of the present invention;
FIG. 2 is a schematic diagram of a dynamic modeling system for a complex system in an industrial process provided by an embodiment of the present invention;
FIG. 3 is a graph of the effect of the proposed method on verifying the effectiveness of the MG system;
FIG. 4 is a graph of the effect of fitting the methods to nonlinear system identification data;
FIG. 5 is a graph of the effect of each method on the fit of butane concentration data at the bottom of the debutanizer;
FIG. 6 is a graph of the effect of the proposed process on the effectiveness of debutanizer bottoms butane concentration data.
Detailed Description
The present invention will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Embodiment one:
The embodiment provides a precise prediction method for a deep recursion random configuration network industrial process, which is applied to a complex nonlinear system dynamic modeling scene. In this example, the method is used to predict product concentration during a chemical reaction.
Data preprocessing: and collecting real-time operation data of the reactor, including parameters such as temperature, pressure, raw material addition amount and the like, and carrying out normalization treatment.
Initializing a network: initial parameters of the network are set, including the number of pool neurons, learning rate, etc.
Randomly configuring weights and biases: and initializing network weights and offsets by a random method, and constructing a pool candidate neuron.
Selecting candidate neurons: and optimizing and selecting candidate neurons which are most suitable for the current data characteristics by using a genetic algorithm according to the real-time data and the system performance requirements.
Increasing optimal candidate neurons: the selected candidate neurons are added to the reservoir to form a new network structure.
Calculating an output weight value: and calculating the output weight of the network by using a least square method.
Error calculation and feedback: and calculating errors according to the real-time product concentration data and the network predicted value, and dynamically adjusting the network structure and parameters.
And (3) continuously repeating the process until the error of the network output reaches an acceptable range, and completing the dynamic modeling of the chemical reaction process.
Embodiment two:
In another embodiment, the accurate prediction method is used for power system load prediction.
The specific implementation steps are as follows:
Data preprocessing: and collecting historical load data of the power system, and taking the influence of factors such as weather, holidays and the like on the power load into consideration to perform data cleaning and preprocessing.
Network initialization: the number of network layers and the number of neurons per layer are set according to the characteristics of the power system load.
Randomly configuring weights and biases: and generating initial weights and biases of the network by adopting a randomization method.
Screening candidate neurons: a sliding window method is used to dynamically select the appropriate candidate neurons based on past load data.
Determining an optimal candidate neuron: candidate neurons capable of maximizing prediction accuracy are screened out through a simulated annealing algorithm.
Calculating an output weight value: and optimizing the output weight according to the target output by adopting a gradient descent method.
Error calculation and adjustment: and carrying out error analysis according to the deviation of the prediction result and the actual load, and gradually optimizing the network structure.
By updating and adjusting the output weight in real time, the model can realize accurate prediction of the power load, thereby improving the efficiency and safety of power grid operation.
The invention is mainly aimed at improving the problems and defects of the following prior art, and realizes remarkable technical progress:
Prediction accuracy and adaptivity are insufficient: existing industrial prediction techniques may suffer from inadequate prediction accuracy and adaptability when processing dynamic, complex industrial data. The invention obviously improves the prediction precision and the adaptability by adopting the deep recursion random configuration network.
Randomness and uncertainty of weights and bias configurations: in a conventional randomly configured network, the random configuration of weights and offsets may lead to unstable network performance. The invention reduces the uncertainty caused by such randomness by supervised selection of candidate neurons and optimization of reservoir structures.
Limitations of dynamic system modeling: the prior art may have limitations in modeling dynamic systems in industrial processes. The invention enhances the modeling capability of the model to the dynamic system by utilizing the projection algorithm updated by the real-time data.
The invention solves the technical effects and remarkable technical progress brought by the prior art problems:
The present invention significantly improves the accuracy of predictions through deep recursive stochastic configuration of the network and well-designed initialization steps, which is critical for optimization and control of complex industrial processes. The invention enables the network to adapt to new data and changes through the projection algorithm of real-time data update, and provides higher adaptability. By supervised selection of optimal candidate neurons, and automatic tuning of reservoir and network layers, the present invention reduces the complexity of model training and tuning. The invention improves the depth and complexity processing capacity of the network by repeatedly adjusting and optimizing the reserve pool neurons in the multi-layer network structure. Due to high precision and adaptability, the invention is suitable for various complex industrial scenes, and improves practicality and universality.
Aiming at the problems existing in the prior art, the invention provides a method and a system for accurately predicting a deep recursive random configuration network industrial process, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the method provided by the invention comprises the following specific steps:
And step1, preprocessing data.
And 2, initializing a deep recursion random configuration network.
Setting the maximum layer number as S, and the maximum allowable number of neurons in each layer of reserve poolInitial neuron number n=0, maximum expected output error tolerance value epsilon, maximum candidate neuron generation number G max, input weight distribution parameter γ= { λ min,λmin+Δλ,...,λmax }, model output error initialization: e 0 = T.
Step 3, randomly selecting an input weight from gamma= { lambda min,λmin+Δλ,...,λmax Pool connection weights/>And bias/>Substitution of activation function/>(First layer) or/>(Layer j) obtaining G max group candidate neurons
Step 4, settingSubstituting G max sets of candidate neurons into the inequality constraint of the random configuration algorithm:
Wherein the method comprises the steps of Is the output error of all pool neurons of the previous layers (N sum = N in the case of single layer structure, and:/>, in the case of layer j)Candidate neurons satisfying the inequality constraint are screened out.
Step 5, defining a group of variables: substituting candidate neurons meeting inequality constraint conditions, when/> Selecting candidate neuron with maximum xi N+1, and reservingAnd/>Carrying out the next step; otherwise, λ=λ+Δλ, and when λ < λ max, the process returns to step 3 to continue execution. If neither of the above conditions is satisfied, let τ e (0, 1-r), r=r+τ, λ=λ min, and then return to step 3 to continue execution. Adding the optimal candidate neuron into a reserve pool to obtain the latest connection weight and bias/>And/>
Step 6, updating the neural network model, and obtaining the output weight of the model by using a least square method through the output of the reserve pool and the target outputFinally updating the model output error/>And pool neuron number n=n+1.
Step 7, when the root mean square error of the model output is larger than or equal to e 0||2, andThen return to step 3 for execution; if ||e 0||2 is not less than ε, and/>Then the next layer is entered for continued training, the pool neuron number n=0 and the pool Chi Cengshu j=j+1 are updated, and the process returns to step 3. If ||e 0||2 < ε, or j=S and/>The training of the deep network model is finished.
Step 8, based on projection algorithm, training the output weight according to the real-time dataUpdating on line, calculating real-time output and analyzing the generalization performance of the model.
As shown in fig. 2, a dynamic modeling system for a complex system in an industrial process according to an embodiment of the present invention includes:
The preprocessing module is used for preprocessing data;
the initialization module is used for initializing the deep recursion random configuration network;
Constructing a neuron module, which is used for randomly configuring weights and biases and constructing candidate neurons of the reserve pool by adopting a brand new method;
The candidate neuron selection module is used for supervised selection of candidate neurons according to inequality constraint conditions;
an optimal neuron determination module that determines optimal candidate neurons to be added to the reservoir;
The output weight calculation module is used for calculating the current layer network output weight by combining the target output and the reserve pool output;
and the online updating module is used for online updating the trained output weight according to the real-time data based on the projection algorithm and calculating real-time output.
Preferably, the method for accurately predicting the deep recursion random configuration network industrial process provided by the embodiment of the invention comprises the following steps:
S1, preprocessing data;
S2, initializing a deep recursion random configuration network;
S3, randomly configuring weights and biases, and constructing candidate neurons of the reserve pool by adopting a brand new method;
s4, selecting candidate neurons in a supervision mode according to inequality constraint conditions;
S5, determining optimal candidate neurons, and adding the optimal candidate neurons into a reserve pool;
s6, calculating the current layer network output weight by combining the target output and the reserve pool output;
s7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3-S6 until the error requirement is met or the maximum allowable number of neurons of the reserve pool of the current layer is reached;
S8, when the error requirement is not met, entering the next layer, and sequentially configuring the storage pool neurons of each layer according to S3-S7 until the error requirement is met or the maximum layer number requirement is met;
And S9, based on a projection algorithm, carrying out online updating on the trained output weight according to the real-time data, and calculating real-time output.
S1 specifically comprises: given a set of timing data, input samples: { u (1), u (2),..u (n max)}＝{(y(1),u(1)),(y(2),u(2)),...,(y(nmax),u(nmax)) }, target output: t= { T (2), T (3), T (n max +1) }, where y, u are the system output and the controlled input, respectively, the present invention predicts the next moment output y (n+1) only by the current moment input (y (n), u (n)), n max is the number of samples. Setting the maximum layer number as S, and the maximum allowable number of neurons in each layer of reserve poolThe maximum expected output error tolerance epsilon, the maximum candidate neuron generation number G max, and the input weight distribution parameter gamma= { lambda min,λmin+Δλ,...,λmax }.
S2 specifically comprises: initializing an output error vector e 0 =t, and outputting an error scaling factor 0 < r < 1 by a model, wherein the network model of different layers is as follows:
…
Where j=1, 2,.., And/>Is the input weight of the j layer, the reserve pool is connected with the weight and the output weight,/>Is the bias, X (j) (n) is the pool state at time n, g is the activation function, X (j)＝[x(j)(1),x(j)(2),...x(j) (n) ] is the pool state matrix, and Y (j) is the network output of layer j.
S3 specifically comprises: randomly selecting an input weight matrix from gamma= { lambda min,λmin+Δλ,...,λmax }, wherein the input weight matrix is a matrix of the input weight matrixPool connection weight matrix/>And bias matrix/>Substitution of activation functions(First layer) or(Layer j) obtaining G max group candidate neurons
S4 specifically comprises the following steps: substituting G max sets of candidate neurons into the inequality constraint of the random configuration algorithm:
Wherein the method comprises the steps of Is the output error of all pool neurons of each layer, when the single-layer structure is adopted: n sum = N, layer j: /(I)M is the output dimension, non-negative real sequence/>Satisfy/>And/>Candidate neurons satisfying the inequality constraint are screened out.
S5 specifically comprises the following steps: defining a set of variables: Candidate neurons satisfying the inequality constraint and maximizing ζ N+1 are selected as optimal pool candidate neurons.
S6 specifically comprises the following steps: adding the optimal reserve pool candidate neurons into a network model, and directly calculating the output weight of the neural network model according to the target outputCalculating the output root mean square error of the neural network model, taking the root mean square error as a loss function, and updating the output error/>, of the modelAnd pool neuron number n=n+1.
S7 specifically comprises the following steps: if the output root mean square error e 0||2 of the current network model is greater than the maximum expected output error tolerance epsilon and the pool neuron number N is less than the pool neuron maximum toleranceRepeating S3-S6; if the output root mean square error of the current network model, e 0||2, is greater than the maximum expected output error tolerance epsilon and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>Then the next layer is entered for continuous training, the number of the neurons in the reserve pool is updated to be N=0, the number of the neurons in the reserve pool is updated to be Chi Cengshu j=j+1, and the steps S3 to S6 are repeated; if the root mean square error of the current network model output i e 0||2 is less than the maximum expected output error tolerance epsilon, or j=s and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>And after the training is finished, obtaining a depth network model meeting the constraint condition of the random configuration theory.
S9 specifically comprises: based on a projection algorithm, online updating is carried out on the trained output weight according to real-time data, and real-time output is calculated:
where 0 < a < 1, c > 0 is two constants, g (n) is the reservoir output at time n, and e (n) is the output error before correction.
S3-S6 are key points of the invention, S3 links input data with weight parameters, weight distribution can be adaptively adjusted, error convergence is accelerated, and the data-dependent parameter selection method can effectively improve training speed and prediction accuracy of a model.
S3, adopting a brand new reserve pool connection weight construction method, and for the (n+1) th newly added neuron, only reserving the network connection weight of the newly added neuron to the original neuronAnd its own network connection weightThe connection weight of the original neuron to the newly added neuron is set to be zero/>Namely, no connection:
Wherein the method comprises the steps of Representing the connection weight of the ith neuron to the jth neuron,/>The construction method can ensure that the output states of the neurons of the first N reserve pools are not changed when each neuron is added in the reserve pools, and only the newly added neurons which reduce the training errors of the model are required to be configured, so that the global approximation property of the network model is ensured.
The principle involved in the inequality constraint of the random configuration algorithm mentioned in S4 is as follows:
it is assumed that the vector space Γ over the L 2 space is dense, while There is 0 < |g the I is less than b g,Given 0 < r < 1 and non-negative real sequence/>Wherein/>The following formula is given:
if the basis functions are random Constructed output weights/>The method meets the following conditions:
And satisfies the inequality constraint as follows:
Then Where T is the target output value,/>Is a model predicted output value with N sum +1 pool neurons. I.e., the constructed deep neural network model has global approximation properties.
S5, in order to find the candidate neuron which enables the model training error to decrease most rapidly, and further, the convergence speed is increased.
S6 calculating the output weightThe specific principle of (2) is as follows:
it is assumed that the vector space Γ over the L 2 space is dense, while Presence/>Given 0 < r < 1 and non-negative real sequence/>Wherein/>The following formula is given:
if the basis functions are random Constructed output weights/>The method meets the following conditions:
And satisfies the inequality constraint as follows:
Then And the optimal output weight value is obtained.
Taking a Mackey-Glass (MG) chaotic system and a nonlinear dynamic system as examples,
Example 1 Mackey-Glass (MG) chaotic system is a very typical chaotic time series, generated by the following differential equation with time lags:
Where τ > 16.8, the entire sequence is chaotic, periodic, non-converging and divergent.
Example 2: system identification is an important element of modern methodology and signal processing, and the dynamics of the system are believed to manifest in varying input and output data. While the actual system is mostly nonlinear, nonlinear system identification becomes an important and complex problem. Given a nonlinear system in which training phase u (n) =1.05×sin (n/45), the test phase mathematical model is as follows:
y(n+1)＝0.72y(n)+0.025y(n-1)u(n-1)+0.01u2(n-2)+0.2u(n-3)
Example 3: the debutanizer process is an important component of a desulfurization and naphtha separation device in the petroleum refining production process, and the process realizes the prediction of the butane concentration y at the bottom through 7 auxiliary variables, namely, the tower top temperature u 1, the tower top pressure u 2, the tower top reflux amount u 3, the tower top product outflow amount u 4, the layer 6 tower plate temperature u 5, the tower bottom temperature 1u 6 and the tower bottom temperature 2u 7, and a nonlinear model can be described as follows:
y(n)＝f(u1(n),u2(n),u3(n),u4(n),u5(n),u5(n-1),
u5(n-2),u5(n-3),(u6(n)+u7(n))/2,
y(n-1),y(n-2),y(n-3),y(n-4))
considering order unknowns, we reset the inputs of the above model, i.e.
y(n)＝f(u1(n),u2(n),u3(n),u4(n),u5(n),u6(n),u7(n),y(n-1))
In the experiment, ESN is selected in the embodiment of the invention, ESN (DeepESN) of a two-layer structure, ESN (DeepESN) of a three-layer structure, and RSCN is compared with the method (RSCN (DeepRSCN) of a two-layer structure and RSCN (DeepRSCN) of a three-layer structure) of the invention. And selecting a standard root mean square error (NRMSE) as an evaluation index of the network prediction performance.
Example 1 Mackey-Glass (MG) chaotic System
And (3) analyzing experimental results, wherein the training error and the testing error of the deep recursion random configuration network are superior to those of a single-layer model and a traditional echo state network model, and the prediction performance of the model is gradually enhanced along with the increase of the layer number. The lower graph is a comparison graph of training time of each model under different pool scales, and the training speed which is still the highest in the depth model can be seen from the graph 3, so that the effectiveness of the method provided by the invention is verified.
Example 2: nonlinear system identification
Fig. 4 is a graph of the fitting effect of each method, the fitting effect of the deep recursive random configuration network is still the best, and by combining the results in the table, we can see that DeepRSCN performs better than other models on training and test sets, and further verifies the effectiveness of the method in dynamic nonlinear identification.
Example 3: soft measurement of butane concentration at bottom of debutanizer
Fig. 5 shows the prediction effect of each neural network model, from which it can be seen that DeepRSCN has the minimum prediction error, the fitting degree of the model output and the target value is higher, and fig. 6 shows the comparison result of the training time of each model under different reservoir scales, and the depth version has higher training speed, so that the effectiveness of the method in the industrial process parameter soft measurement is verified.
The application of the accurate prediction method for the deep recursive random configuration network industrial process can bring about remarkable technical progress in industry, including higher prediction precision, faster training speed, better robustness and higher resource utilization efficiency. The following are two specific application examples.
Application example one: industrial robot fault prediction
In the field of industrial robots, fault prediction is an important task. The accurate prediction method for the deep recursion random configuration network industrial process can improve the precision and the instantaneity of fault prediction, thereby avoiding unexpected shutdown and improving the production efficiency.
1. Data preprocessing: various sensor data of the industrial robot, such as temperature, pressure, vibration and the like, are collected. The data is normalized or normalized for model training.
2. Deep recursive random configuration network initialization: and (3) carrying out network initialization, pool neuron configuration, optimal neuron determination, network output weight calculation and online updating of real-time output according to the steps S2-S9.
3. And (3) fault prediction: and calculating real-time output of the industrial robot according to the trained model, and predicting whether a fault occurs.
The method has the advantages of remarkably improving the instantaneity and the prediction precision, being capable of finding faults earlier, avoiding unexpected shutdown and improving the production efficiency.
Application example two: smart grid load prediction
In the smart grid field, load prediction is a critical task. The accurate prediction method for the deep recursion random configuration network industrial process can improve the accuracy and the instantaneity of load prediction, so that more effective power scheduling is realized, and the utilization efficiency of power resources is improved.
1. Data preprocessing: load data of the smart grid and various factors affecting the load, such as weather conditions, holiday information and the like, are collected. The data is normalized or normalized for model training.
2. Deep recursive random configuration network initialization: and (3) carrying out network initialization, pool neuron configuration, optimal neuron determination, network output weight calculation and online updating of real-time output according to the steps S2-S9.
3. Load prediction: and calculating real-time output of the intelligent power grid according to the trained model, and predicting future load conditions.
The method has the advantages of remarkably improving the instantaneity and the prediction precision, realizing more effective power scheduling and improving the utilization efficiency of power resources. It should be noted that the embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or special purpose design hardware. Those of ordinary skill in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such as provided on a carrier medium such as a magnetic disk, CD or DVD-ROM, a programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device of the present invention and its modules may be implemented by hardware circuitry, such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., as well as software executed by various types of processors, or by a combination of the above hardware circuitry and software, such as firmware.
The foregoing is merely illustrative of specific embodiments of the present invention, and the scope of the invention is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present invention will be apparent to those skilled in the art within the scope of the present invention.

Claims (4)

1. The accurate prediction method for the deep recursion random configuration network industrial process is characterized in that the accurate prediction method based on the deep recursion random configuration network is adopted, and dynamic system modeling with high adaptability and high prediction precision is realized by data preprocessing, network hierarchy initialization, random configuration weight and bias, selection and optimization of reserve pool neurons in combination with inequality constraint conditions and a projection algorithm updated by real-time data;
The accurate prediction method for the deep recursion random configuration network industrial process comprises the following steps:
S1, preprocessing data;
S2, initializing a deep recursion random configuration network;
S3, randomly configuring weights and biases, and constructing candidate neurons of the reserve pool by adopting a brand new method;
s4, selecting candidate neurons in a supervision mode according to inequality constraint conditions;
S5, determining optimal candidate neurons, and adding the optimal candidate neurons into a reserve pool;
s6, calculating the current layer network output weight by combining the target output and the reserve pool output;
s7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3-S6 until the error requirement is met or the maximum allowable number of neurons of the reserve pool of the current layer is reached;
S8, when the error requirement is not met, entering the next layer, and sequentially configuring the storage pool neurons of each layer according to S3-S7 until the error requirement is met or the maximum layer number requirement is met;
S9, based on a projection algorithm, online updating is carried out on the trained output weight according to real-time data, and real-time output is calculated;
s1 specifically comprises: given a set of timing data, input samples: { (y (1), u (1)), (y (2), u (2)), (y (n max),u(nmax)) } target output: t= { T (2), T (3), T (n max +1) }, where y, u are the system output and the controlled input, respectively, and the current time input (y (n), u (n)) predicts the next time output y (n+1), and n max is the number of samples; setting the maximum layer number as S, and the maximum allowable number of neurons in each layer of reserve pool Maximum expected output error tolerance value epsilon, maximum candidate neuron generation number G max, input weight distribution parameter gamma= { lambda min,λmin+Δλ,...,λmax };
S2 specifically comprises: initializing an output error vector e 0 =t, and outputting an error scaling factor 0 < r < 1 by a model, wherein the network model of different layers is as follows:
…
Where j=1, 2,.., And/>Is the input weight of the j layer, the reserve pool is connected with the weight and the output weight,/>Is the bias of the j-th layer, X (j) (n) is the pool state of the j-th layer at time n, g is the activation function, X (j)＝[x(j)(1),x(j)(2),...x(j) (n) is the pool state matrix, and Y (j) is the network output of the j-th layer;
S3 specifically comprises: randomly selecting the weight value input into the j-th layer from gamma= { lambda min,λmin+Δλ,...,λmax }, wherein the weight value is selected from gamma = { lambda min,λmin+Δλ,...,λmax }, and the weight value is input into the j-th layer The reserve pool is connected with the weight value/>, of the j th layerAnd bias of the j-th layer/>Substitution of activation functionsOr (b)Obtaining G max group candidate neurons
S4 specifically comprises the following steps: substituting G max sets of candidate neurons into the inequality constraint of the random configuration algorithm:
Wherein the method comprises the steps of Is the output error of all pool neurons of each layer, when the single-layer structure is adopted: n sum = N, layer j: m is the output dimension, non-negative real sequence/> Satisfy/>And is also provided withScreening candidate neurons meeting inequality constraint;
s5 specifically comprises the following steps: defining a set of variables:
screening candidate neurons which meet inequality constraint and maximize xi N+1 from the candidate neurons as optimal reserve pool candidate neurons;
s6 specifically comprises the following steps: adding the optimal reserve pool candidate neurons into a network model, and directly calculating the output weight of the neural network model according to the target output Calculating the output root mean square error of the neural network model, taking the root mean square error as a loss function, and updating the output error/>, of the modelAnd pool neuron number n=n+1.
2. The method for accurately predicting a deep recursive random configuration network industrial process of claim 1, wherein S7 specifically comprises: if the output root mean square error e 0||2 of the current network model is greater than the maximum expected output error tolerance epsilon and the pool neuron number N is less than the pool neuron maximum toleranceRepeating S3-S6; if the output root mean square error of the current network model, e 0||2, is greater than the maximum expected output error tolerance epsilon and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>Then the next layer is entered for continuous training, the number of the neurons in the reserve pool is updated to be N=0, the number of the neurons in the reserve pool is updated to be Chi Cengshu j=j+1, and the steps S3 to S6 are repeated; if the root mean square error of the current network model output i e 0||2 is less than the maximum expected output error tolerance epsilon, or j=s and the current layer pool neuron number N is equal to the pool neuron maximum tolerance/>And after the training is finished, obtaining a depth network model meeting the constraint condition of the random configuration theory.
3. The method for accurately predicting a deep recursive random configuration network industrial process of claim 1, wherein S9 specifically comprises: based on a projection algorithm, online updating is carried out on the trained output weight according to real-time data, and real-time output is calculated:
where 0 < a < 1, c > 0 is two constants, g (n) is the reservoir output at time n, and e (n) is the output error before correction.
4.A dynamic modeling system applying the deep recursive random configuration network industrial process precision prediction method of any one of claims 1 to 3, comprising:
The preprocessing module is used for preprocessing data;
the initialization module is used for initializing the deep recursion random configuration network;
Constructing a neuron module, which is used for randomly configuring weights and biases and constructing candidate neurons of the reserve pool by adopting a brand new method;
The candidate neuron selection module is used for supervised selection of candidate neurons according to inequality constraint conditions;
an optimal neuron determination module that determines optimal candidate neurons to be added to the reservoir;
The output weight calculation module is used for calculating the current layer network output weight by combining the target output and the reserve pool output;
and the online updating module is used for online updating the trained output weight according to the real-time data based on the projection algorithm and calculating real-time output.
CN202311743288.7A 2023-12-18 2023-12-18 Accurate prediction method and system for deep recursion random configuration network industrial process Active CN117709536B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311743288.7A CN117709536B (en) 2023-12-18 2023-12-18 Accurate prediction method and system for deep recursion random configuration network industrial process

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311743288.7A CN117709536B (en) 2023-12-18 2023-12-18 Accurate prediction method and system for deep recursion random configuration network industrial process

Publications (2)

Publication Number Publication Date
CN117709536A CN117709536A (en) 2024-03-15
CN117709536B true CN117709536B (en) 2024-05-14

Family

ID=90154994

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311743288.7A Active CN117709536B (en) 2023-12-18 2023-12-18 Accurate prediction method and system for deep recursion random configuration network industrial process

Country Status (1)

Country Link
CN (1) CN117709536B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115081484A (en) * 2022-07-04 2022-09-20 南京航空航天大学 Aircraft engine sensor fault diagnosis method based on CRJ-OSELM algorithm
WO2023115596A1 (en) * 2021-12-21 2023-06-29 浙江工业大学台州研究院 Truss stress prediction and weight lightening method based on transfer learning fusion model
WO2023168916A1 (en) * 2022-03-08 2023-09-14 太原理工大学 Neural network model optimization method based on stainless steel ultra-thin strip annealing process

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170091615A1 (en) * 2015-09-28 2017-03-30 Siemens Aktiengesellschaft System and method for predicting power plant operational parameters utilizing artificial neural network deep learning methodologies
US11354747B2 (en) * 2016-04-16 2022-06-07 Overbond Ltd. Real-time predictive analytics engine
EP3620983B1 (en) * 2018-09-05 2023-10-25 Sartorius Stedim Data Analytics AB Computer-implemented method, computer program product and system for data analysis

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023115596A1 (en) * 2021-12-21 2023-06-29 浙江工业大学台州研究院 Truss stress prediction and weight lightening method based on transfer learning fusion model
WO2023168916A1 (en) * 2022-03-08 2023-09-14 太原理工大学 Neural network model optimization method based on stainless steel ultra-thin strip annealing process
CN115081484A (en) * 2022-07-04 2022-09-20 南京航空航天大学 Aircraft engine sensor fault diagnosis method based on CRJ-OSELM algorithm

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于压缩感知的回声状态神经网络在时间序列预测中的应用;李莉;於志勇;黄昉菀;;软件导刊;20200415(第04期);第15-19页 *

Also Published As

Publication number Publication date
CN117709536A (en) 2024-03-15

Similar Documents

Publication Publication Date Title
Polo et al. Failure mode prediction and energy forecasting of PV plants to assist dynamic maintenance tasks by ANN based models
Garg Reliability, availability and maintainability analysis of industrial systems using PSO and fuzzy methodology
CN108764539B (en) Upstream and downstream water level prediction method for cascade power station
Attarzadeh et al. Proposing an enhanced artificial neural network prediction model to improve the accuracy in software effort estimation
CN111680786B (en) Time sequence prediction method based on improved weight gating unit
Yang et al. A novel self-constructing radial basis function neural-fuzzy system
CN101782743A (en) Neural network modeling method and system
CN113065703A (en) Time series prediction method combining multiple models
CN111950810A (en) Multivariable time sequence prediction method and device based on self-evolution pre-training
Suryo et al. Improved time series prediction using LSTM neural network for smart agriculture application
CN116526473A (en) Particle swarm optimization LSTM-based electrothermal load prediction method
Jichang et al. Water quality prediction model based on GRU hybrid network
Pranolo et al. Optimized three deep learning models based-pso hyperparameters for beijing pm2. 5 prediction
JPWO2020021687A1 (en) Water treatment plant
Liu et al. A robust reliability prediction method using weighted least square support vector machine equipped with chaos modified particle swarm optimization and online correcting strategy
CN117235595A (en) LSTM-Attention-based sewage treatment dosing prediction method and system
Abdollahzadeh et al. Design of a novel robust adaptive neural network‐based fractional‐order proportional‐integrated‐derivative controller on DC/DC Boost converter
Zhou et al. Deep learning-based intelligent multilevel predictive maintenance framework considering comprehensive cost
CN117709536B (en) Accurate prediction method and system for deep recursion random configuration network industrial process
CN118017482A (en) Flexible climbing capacity demand analysis method based on prediction error feature extraction
CN116992935A (en) Neural network model training method and system based on improved random configuration algorithm
Guan et al. Bounded error modeling using interval neural networks with parameter optimization
Fang et al. Twinning quality sensors in wastewater treatment process via optimized echo state network-based soft sensors
CN115907200A (en) Concrete dam deformation prediction method, computer equipment and storage medium
CN114897245A (en) Large-scale wind power ultra-short term power prediction error correction method based on long-short term memory neural network

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