WO2022121932A1 - 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 - Google Patents
基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 Download PDFInfo
- Publication number
- WO2022121932A1 WO2022121932A1 PCT/CN2021/136373 CN2021136373W WO2022121932A1 WO 2022121932 A1 WO2022121932 A1 WO 2022121932A1 CN 2021136373 W CN2021136373 W CN 2021136373W WO 2022121932 A1 WO2022121932 A1 WO 2022121932A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- deep learning
- model
- online
- prediction model
- offline
- Prior art date
Links
- 238000013135 deep learning Methods 0.000 title claims abstract description 303
- 238000013277 forecasting method Methods 0.000 title claims abstract description 15
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 11
- 238000003860 storage Methods 0.000 title claims abstract description 6
- 238000012937 correction Methods 0.000 claims abstract description 109
- 210000002569 neuron Anatomy 0.000 claims description 55
- 239000003518 caustics Substances 0.000 claims description 52
- 239000003513 alkali Substances 0.000 claims description 46
- 238000004422 calculation algorithm Methods 0.000 claims description 38
- 238000001514 detection method Methods 0.000 claims description 38
- 238000004519 manufacturing process Methods 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 32
- PNEYBMLMFCGWSK-UHFFFAOYSA-N aluminium oxide Inorganic materials [O-2].[O-2].[O-2].[Al+3].[Al+3] PNEYBMLMFCGWSK-UHFFFAOYSA-N 0.000 claims description 28
- 238000000034 method Methods 0.000 claims description 28
- 238000005259 measurement Methods 0.000 claims description 11
- 238000003556 assay Methods 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 abstract description 5
- 230000008569 process Effects 0.000 description 15
- 230000006870 function Effects 0.000 description 9
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000001704 evaporation Methods 0.000 description 7
- 230000008020 evaporation Effects 0.000 description 7
- 238000005070 sampling Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000005457 optimization Methods 0.000 description 5
- 229910001570 bauxite Inorganic materials 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 238000002360 preparation method Methods 0.000 description 4
- WNROFYMDJYEPJX-UHFFFAOYSA-K aluminium hydroxide Chemical compound [OH-].[OH-].[OH-].[Al+3] WNROFYMDJYEPJX-UHFFFAOYSA-K 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- ANBBXQWFNXMHLD-UHFFFAOYSA-N aluminum;sodium;oxygen(2-) Chemical compound [O-2].[O-2].[Na+].[Al+3] ANBBXQWFNXMHLD-UHFFFAOYSA-N 0.000 description 2
- 210000004027 cell Anatomy 0.000 description 2
- 239000013078 crystal Substances 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000007787 long-term memory Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 239000002994 raw material Substances 0.000 description 2
- 229910001388 sodium aluminate Inorganic materials 0.000 description 2
- 238000004131 Bayer process Methods 0.000 description 1
- 235000008733 Citrus aurantifolia Nutrition 0.000 description 1
- 238000003723 Smelting Methods 0.000 description 1
- 235000011941 Tilia x europaea Nutrition 0.000 description 1
- 229910052782 aluminium Inorganic materials 0.000 description 1
- XAGFODPZIPBFFR-UHFFFAOYSA-N aluminium Chemical compound [Al] XAGFODPZIPBFFR-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 238000000227 grinding Methods 0.000 description 1
- 239000004571 lime Substances 0.000 description 1
- 238000002844 melting Methods 0.000 description 1
- 230000008018 melting Effects 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 229910052751 metal Inorganic materials 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 239000012452 mother liquor Substances 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 238000000746 purification Methods 0.000 description 1
- 239000011819 refractory material Substances 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000003756 stirring Methods 0.000 description 1
Images
Classifications
-
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- 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
- G06N3/048—Activation functions
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P80/00—Climate change mitigation technologies for sector-wide applications
- Y02P80/10—Efficient use of energy, e.g. using compressed air or pressurized fluid as energy carrier
Definitions
- the invention belongs to the technical field of industrial artificial intelligence, and relates to an intelligent forecasting method, device, equipment and storage medium for complex industrial systems based on adaptive deep learning.
- the forecasting model of production indicators and key process parameters is required to give forecast values in the decision-making time period. This requires that the training data set of the deep learning forecast model should not be too large, and the training algorithm should not take too long. Coupled with the complexity of the manufacturing process, the production indicators, key process parameters and related production process input and output variables are complex dynamic systems.
- the dynamic system often has strong nonlinearity, strong coupling of multiple variables, unknown order of model structure and input and output variables, or even changes, changes in production boundary conditions such as raw materials, material flow, information flow, and energy flow in the production process.
- the interaction of the dynamic system makes the characteristics of the dynamic system change unknown with the production time, resulting in the input and output data of the system in a changing, open and uncertain information space, which makes the existing deep learning technology of complete information space. It cannot be applied to the above-mentioned complex industrial dynamic system to establish the forecasting model of the system.
- the present invention aims to solve one of the technical problems in the related art at least to a certain extent.
- the technical scheme of the present invention is as follows:
- An intelligent forecasting method for complex industrial systems based on adaptive deep learning comprising the following steps:
- the online deep learning prediction model is used for real-time prediction of the parameters of the complex industrial system.
- the establishment of a dynamic model of a complex industrial system includes: determining input variables and output variables of the dynamic model, and the output variables are predicted variables; and using the dynamic model to establish an offline deep learning forecast
- the model includes: using LSTM to establish the offline deep learning prediction model, using the input variable of the dynamic model as the input of the LSTM, using the output data of the dynamic model as the label data, using an offline training algorithm, according to the label.
- the error between the data and the output of the offline deep learning prediction model determines the number of neurons, the number of unit nodes, the number of network layers, and the weight parameters and bias parameters of each layer of the LSTM; the use of the offline deep learning
- the prediction model to establish an online deep learning prediction model includes: using LSTM to establish the online deep learning prediction model, and the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the online deep learning prediction model are all the same as the number of all network layers.
- the offline deep learning prediction model is the same, and the weight parameters and bias parameters of each layer of the offline deep learning prediction model are used as the initial value of the weight parameter and the initial value of the bias parameter of the corresponding layer of the online deep learning prediction model.
- a deep learning correction model with the same structure as the online deep learning prediction model includes: using LSTM to establish the deep learning correction model, the input of a single neuron of the deep learning correction model, the number of neurons, the number of unit nodes and The number of network layers is the same as the online deep learning prediction model.
- the use of the deep learning correction model to correct the online deep learning prediction model includes: when a preset condition is met, using the weight parameters and bias parameters of each layer of the deep learning correction model The weight parameters and bias parameters of the corresponding layers of the online deep learning forecast model are replaced; wherein, the historical data input by the deep learning correction model is more than the historical data input by the online deep learning forecast model.
- the online correction of the weight parameters and bias parameters of the last layer of the online deep learning prediction model specifically, the online correction of part of the weight parameters and the last layer of the online deep learning prediction model. Partial bias parameter.
- the complex industrial system is an alumina preparation system
- the online deep learning prediction model is used to perform real-time prediction on the detection error of the caustic alkali concentration of the alumina preparation system; the caustic alkali concentration detection
- the error is the difference between the assay value of the caustic alkali concentration and the measured value of the caustic alkali concentration online detection instrument.
- An intelligent forecasting device for complex industrial systems based on adaptive deep learning comprising:
- Dynamic model modeling module used to establish dynamic models of complex industrial systems
- An offline deep learning forecast model modeling module used for establishing an offline deep learning forecast model by using the dynamic model
- an online deep learning forecast model modeling module used for using the offline deep learning forecast model to establish an online deep learning forecast model
- a deep learning correction model modeling module used for establishing a deep learning correction model using the same structure as the online deep learning prediction model
- a self-correction module for correcting the online deep learning prediction model by using the deep learning correction model
- the online deep learning prediction model is used for real-time prediction of the parameters of the complex industrial system.
- the dynamic model modeling module determines the input variables and output variables of the dynamic model, and the output variables are predicted variables; the offline deep learning prediction model modeling module adopts LSTM to establish the offline A deep learning prediction model, using the input variable of the dynamic model as the input of the LSTM, using the output data of the dynamic model as label data, using an offline training algorithm, according to the label data and the offline deep learning prediction model
- the error between the outputs determines the number of neurons, the number of unit nodes, the number of network layers, and the weight parameters and bias parameters of each layer of the LSTM;
- the online deep learning prediction model modeling module uses LSTM to establish the online depth Learning forecasting model, the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the online deep learning forecasting model are all the same as the offline deep learning forecasting model.
- the weight parameters and bias parameters of each layer are used as the initial value of the weight parameter and the initial value of the bias parameter of the corresponding layer of the online deep learning prediction model, and an online training algorithm is adopted. Predict the error between the outputs of the model, correct the weight parameters and bias parameters of the last layer of the online deep learning prediction model online; the deep learning correction model modeling module adopts LSTM to establish the deep learning correction model, so The input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the deep learning correction model are all the same as the online deep learning prediction model.
- the error between the model outputs is corrected in real time for the weight parameters and bias parameters of each layer of the deep learning correction model; when the self-correction module meets the preset conditions, the The weight parameter and the bias parameter replace the weight parameter and the bias parameter of the corresponding layer of the online deep learning prediction model; wherein, the historical data input by the deep learning correction model is larger than the historical data input by the online deep learning prediction model. Lots of data.
- the online correction of the weight parameters and bias parameters of the last layer of the online deep learning prediction model specifically, the online correction of part of the weight parameters and the last layer of the online deep learning prediction model. Partial bias parameter.
- the complex industrial system is an alumina preparation system
- the online deep learning prediction model is used to perform real-time prediction on the detection error of the caustic alkali concentration of the alumina preparation system; the caustic alkali concentration detection
- the error is the difference between the assay value of the caustic alkali concentration and the measured value of the caustic alkali concentration online detection instrument.
- An intelligent forecasting device for complex industrial systems based on adaptive deep learning for implementing the above-mentioned intelligent forecasting method comprising: terminal-side sub-devices, edge-side sub-devices, and cloud-side sub-devices;
- the end-side sub-device is used to collect input data and output data of the complex industrial system
- the edge side sub-device uses the online deep learning prediction model to perform real-time prediction on the parameters of the complex industrial system
- the cloud side sub-device is used to train the deep learning correction model, and realize the correction of the online deep learning prediction model by the deep learning correction model.
- a computer-readable storage medium storing a computer program, when the program is executed by a processor, implements the above-mentioned intelligent forecasting method for a complex industrial system.
- the present invention establishes a mechanism including an offline deep learning forecasting model, an online deep learning forecasting model, a deep learning correction model and self-correction, and realizes accurate real-time forecasting of complex industrial systems. forecast.
- Fig. 1 is the realization flow chart of the intelligent forecasting method of complex industrial system according to the embodiment of the present invention
- Fig. 2 is the realization flow chart of the intelligent prediction method of caustic alkali concentration detection error of one embodiment of the present invention
- Figure 3 shows the forecast error of the online deep learning forecast model when the input data time series window length takes different values
- FIG. 4 is a schematic structural diagram of an intelligent forecasting device for a complex industrial system according to an embodiment of the present invention.
- FIG. 5 is a schematic structural diagram of an intelligent forecasting device for a complex industrial system according to an embodiment of the present invention.
- Fig. 1 is the realization flow chart of the complex industrial system intelligent forecasting method of the embodiment of the present invention, the method comprises the following steps:
- the production index or key process parameter that needs to be predicted is the output variable of the dynamic model of the industrial system
- the input and output of the industrial process that affects the output variable are the input variables of the dynamic model
- the The output historical data of the dynamic model and the historical prediction error data are used as the input variables of the dynamic model
- the unknown constant n is used to represent the order of unknown changes of the input and output variables of the dynamic system.
- the dynamic model of the industrial system is represented by the following equation:
- s(k) f(s(k-1),...,s(kn),y i (k),...,y i (k-n+1),u i (k),...,u i ( k-n+1), ⁇ s(k-1),..., ⁇ s(kn)) (1)
- f is the nonlinear function of unknown change
- s(k) is the output of the dynamic model at time k
- y i (k) is the ith output of the industrial process at time k
- u i (k) is the industrial process at time k
- the offline deep learning prediction model is established by using LSTM, the input variable of the dynamic model is used as the input of the LSTM, the output data of the dynamic model is used as the label data, and an offline training algorithm is used, according to the label data.
- the error between the data and the output of the offline deep learning prediction model determines the number of neurons, the number of unit nodes, the number of network layers, and the weight parameters and bias parameters of each layer of the LSTM.
- Step S2 includes steps S21 and S22.
- Step S21 is: using the LSTM structure to establish an offline deep learning prediction model, setting the initial network layer number of the LSTM to 1, and determining the number of neurons of the LSTM according to the difference between the label data and the output of the offline deep learning prediction model through a training algorithm. and the number of element nodes.
- x j (k+jn) [s(k+jn-1),y i (k+jn),u i (k+jn), ⁇ s(k+jn-1))] T (2)
- the output data s(k) of the industrial system dynamic model (1) is used as the label data, the input and output data of formula (1) are used to form a large data sample, and the offline training algorithm is used to make the label data and the output of the offline deep learning prediction model match.
- the difference is as small as possible, and the number of neurons n and the number of unit nodes h of the LSTM are determined.
- Step S22 is: fixing the number of neurons and unit nodes of the LSTM, changing the number of network layers of the LSTM, according to the difference between the label data corresponding to the different network layers and the output of the offline deep learning prediction model, Choose the number of network layers for the LSTM.
- the number of single neurons n of the LSTM and the number of LSTM unit nodes h are fixed, and the error between the output of the offline deep learning prediction model and the label data is reduced by increasing the number of network layers of the LSTM, so that the error is as small as possible , determine the number of LSTM network layers and the weight parameters and bias parameters of each layer.
- the online deep learning prediction model is established by using LSTM, and the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the online deep learning prediction model are the same as those of the offline deep learning prediction model.
- the sequence N is used to online correct the weight parameters and bias parameters of the last layer of the online deep learning prediction model to ensure that the online deep learning prediction model completes the prediction algorithm within the determined optimization decision time period.
- N is determined by making the forecast error as small as possible.
- a data set with a time series length of N is used, and a recursive algorithm is used, that is, the time sequence of the input data of the online deep learning forecast model at time k is (k-N+1),...,k; the online depth at time (k+1)
- the time sequence of the input data of the learning forecast model is (k-N+2),...,(k+1).
- Online deep learning forecasting models are used for real-time forecasting of parameters of complex industrial systems.
- LSTM is used to establish the deep learning correction model.
- the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the deep learning correction model are the same as those of the online deep learning prediction model.
- the input data of the model (1) at the current moment and all previous moments is used as the input data of the deep learning correction model, and the ownership value parameters and bias parameters of each layer of the deep learning correction model are trained to obtain the predicted value of the deep learning correction model. and forecast error
- the online deep learning prediction model is adaptively corrected by using the deep learning correction model, and the upper bound of the prediction error interval is set as ⁇ .
- the prediction error of the online deep learning prediction model
- deep learning corrects the prediction error of the model
- the weight parameters and bias parameters of each layer of the deep learning correction model are used to replace the weight parameters and bias parameters of the corresponding layers of the online deep learning prediction model to ensure that the prediction error of the online deep learning prediction model is within the range of the set prediction error. , that is,
- the intelligent forecasting method for complex industrial systems can be used for forecasting errors in detection of caustic alkali concentration in an alumina production system.
- Alumina has the excellent characteristics of high hardness and high melting point. It is often used in smelting metal aluminum and making refractory materials. It is a strategic resource that plays a supporting role in the military industry, aerospace and national economy.
- the main method for producing alumina is the Bayer process.
- the basic process is usually to add crushed bauxite into lime and caustic solution for grinding according to the ratio requirements, and then use the caustic solution to grind at a certain temperature and a certain pressure.
- the sodium aluminate solution is obtained by dissolving bauxite.
- aluminum hydroxide crystals are obtained by cooling, adding crystal seeds, stirring and analyzing, and the precipitated aluminum hydroxide is separated, washed and roasted to obtain alumina.
- the mother liquor (the main component is caustic alkali) after the separation of aluminum hydroxide is re-dissolved new bauxite after the evaporation process, and enters the next cycle.
- the caustic concentration of alumina solution is a key process index in the alumina evaporation process, which is related to the final product quality of alumina. Routine caustic alkali concentration detection relies on manual sampling at a fixed period and then testing to obtain accurate caustic alkali concentration values. However, due to the long sampling interval and testing time, the detection of caustic alkali concentration has a serious lag and cannot achieve evaporation. Optimal control of the operation of the process.
- the input and output data of the system are in a changing, open and uncertain information space, which makes it difficult for the existing deep learning technology with complete information space to be applied to the dynamic system of caustic concentration error prediction in the alumina evaporation process.
- the forecast model of production indicators and key process parameters is required to give forecast values in the decision time period. This requires that the training data set of the deep learning forecast model should not be too large, and the training algorithm should not take too long.
- the embodiment of the present invention establishes a mechanism including an offline deep learning prediction model, an online deep learning prediction model, a deep learning correction model and self-correction, and realizes the realization of alumina production. Produce accurate real-time forecasts of the system.
- Fig. 2 is the realization flow chart of the intelligent prediction method of caustic alkali concentration detection error of one embodiment of the present invention, and this method comprises the following steps:
- S1' Establish a dynamic model of detection error between the assay value of caustic alkali concentration and the measured value of the caustic alkali concentration online detection instrument.
- the input of the detection error dynamic model includes the refractive index and temperature of the alumina solution, and the caustic alkali
- the historical value of the difference between the concentration assay value and the measurement value of the caustic concentration detection instrument is used as the input of the detection error dynamic model.
- the unknown constant n is used to represent the unknown order of the input and output variables of the dynamic system, and the caustic alkali concentration detection error dynamic is established.
- the model is as follows:
- y 1 (k) is the refractive index of the alumina solution at time k
- y 2 (k) is the temperature of the alumina solution at time k
- It is the test value r(k) of the caustic alkali concentration at time k and the measured value of the caustic alkali concentration detection instrument Difference.
- S2' Use the detection error dynamic model to establish an offline deep learning prediction model.
- LSTM is used to establish the offline deep learning prediction model
- the input variable of the detection error dynamic model is used as the input of the LSTM
- the output data of the detection error dynamic model is used as label data
- an offline training algorithm is used
- the number of neurons, the number of unit nodes, the number of network layers, and the weight parameters and bias parameters of each layer of the LSTM are determined.
- Step S2' includes steps S21' and S22'.
- Step S21' is: adopting the LSTM structure to establish an offline deep learning prediction model, setting the initial network layer number of the LSTM to 1, and determining the number of neurons of the LSTM according to the difference between the label data and the output of the offline deep learning prediction model through a training algorithm. number and the number of element nodes.
- x j (k+jn) [y 1 (k+jn),y 2 (k+jn), ⁇ r(k+jn-1)] T (4)
- j 1,...,n; n is the number of neurons.
- the following training algorithm is used to determine the number of neurons n and the number of unit nodes of LSTM
- the objective function of the training algorithm is:
- M represents the amount of training data.
- Forecast values of offline deep learning forecast models is the weighted expression of the nth neuron output h(k):
- h(k) is vector
- W d represents the weight parameter
- W d is vector
- b d represents the bias parameter
- h(k-1) is the output of the (n-1)th neuron
- [h(k-1), x j (k)] T is vector
- W o and b o are the connection weights and biases of the first layer of the neural network
- W o is matrix
- b o is vector
- ⁇ is a sigmoid function
- ⁇ (z) (1+e -z ) -1
- z is an element of the vector [W o ⁇ [h(k-1), x j (k)] T +b o ].
- C(k) is the long-term memory state
- C(k) is vector
- tanh( ) is the hyperbolic tangent function
- c i (k) is the ith element of the vector C(k)
- W f , Wi , and W C are the connection weights of LSTM units, which are all matrix, b f , b i , b C are the LSTM cell biases, all vector.
- Step S22' is: fixing the number of neurons n of the offline deep learning prediction model to 20, and simultaneously setting the number of unit nodes to 20. Fixed to 180, the offline deep learning prediction model can be output by increasing the number of network layers The error between the label data ⁇ r(k) and the label data ⁇ r(k) is as small as possible to determine the number of layers L.
- Forecast values of offline deep learning forecast models is the weighted expression of the output h L (k) of the 20th neuron of the L-th layer LSTM:
- h L (k) is a 180 ⁇ 1 vector, represents the weight parameter, is a 1 ⁇ 180 vector, represents the bias parameter.
- h L (k-1) is the output of the 19th neuron of the L-th layer of LSTM neural network
- h L-1 (k) is the output of the 20th neuron of the L-1th layer of LSTM neural network
- CL (k) is the long-term memory state
- CL (k) is a 180 ⁇ 1 vector
- LSTM cell bias and is a 180 ⁇ 1 vector.
- S3' Use the offline deep learning forecast model to establish an online deep learning forecast model.
- the online deep learning prediction model is established by using LSTM, and the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the online deep learning prediction model are the same as those of the offline deep learning prediction model. , take the weight parameter and bias parameter of each layer of the offline deep learning forecast model as the initial value of the weight parameter and the initial value of the bias parameter of the corresponding layer of the online deep learning forecast model, and correct the online deep learning online Connection weights for the second layer of the forecast model and bias
- the online deep learning forecast model is as follows:
- h 2 (k) is the output of the last neuron of the second layer LSTM unit.
- the window length N of the input data time series of the online deep learning forecast model is determined by traversal.
- the objective function is:
- the online deep learning prediction model for the detection error of caustic alkali concentration at time (k+1) is:
- the online deep learning forecast model at time (k+1) uses the input data of the time series (k-818), (k-817),..., (k+1) with N being 820, and uses the following algorithm to correct the weight parameters and bias parameters Calculate the error prediction value of caustic alkali concentration detection at time (k+1) by formula (21)
- LSTM is used to establish the deep learning correction model, and the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the deep learning correction model are the same as those of the online deep learning prediction model. All input data of the model (3) of the current k moment and all previous moments, namely k,...,2,1 are used as the input data of the deep learning correction model, and the following objective function and training algorithm are used to correct the first step of the deep learning correction model. Ownership value parameters and bias parameters for the first and second layers.
- the objective function is:
- the correction algorithm is as follows.
- the weight parameters and bias parameters of each layer of the deep learning correction model are used to correct the weight parameters and bias parameters of the online deep learning forecast model, so as to ensure that the forecast error of the online deep learning forecast model is within the set value. within the range of the forecast error.
- Table 2 shows the effect of applying the caustic alkali concentration detection error prediction method of the embodiment of the present invention to the evaporation process of an alumina plant in Shanxi.
- the meter measurement value in Table 2 is the online measurement value of the caustic alkali concentration meter
- the compensated meter value is the sum of the online measurement value of the caustic alkali concentration meter and the prediction value output by the online deep learning prediction model of the caustic alkali concentration detection error.
- Table 2 respectively counts the root mean square error RMSE of the instrument measurement value, the compensated instrument value and the caustic alkali concentration assay value, and the pass rate within the error interval specified by the production process.
- the RMSE between the caustic alkali concentration meter measurement value and the assay value can be reduced from 11.25 to 0.50, which is qualified.
- the rate is increased from 10.75% to 99.62%, which creates conditions for realizing the closed-loop operation optimization control of the alumina evaporation process.
- an intelligent forecasting device for complex industrial systems based on adaptive deep learning including: a dynamic model modeling module, an offline deep learning forecasting model modeling module, and an online deep learning forecasting module.
- Model modeling module, deep learning calibration model modeling module and self-calibration module wherein:
- the dynamic model modeling module is used to establish dynamic models of complex industrial systems
- the offline deep learning forecast model modeling module is used to establish an offline deep learning forecast model by using the dynamic model
- the online deep learning forecasting model modeling module is used to establish an online deep learning forecasting model by using the offline deep learning forecasting model
- the deep learning correction model modeling module is used to establish a deep learning correction model using the same structure as the online deep learning prediction model;
- the self-correction module is used to correct the online deep learning prediction model by using the deep learning correction model
- the online deep learning prediction model is used for real-time prediction of the parameters of the complex industrial system.
- the dynamic model modeling module determines input variables and output variables of the dynamic model, and the output variables are predicted variables; the offline deep learning prediction model modeling module uses LSTM to establish the The offline deep learning prediction model, the input variable of the dynamic model is used as the input of the LSTM, the output data of the dynamic model is used as the label data, and an offline training algorithm is adopted.
- the online deep learning prediction model modeling module uses LSTM to establish the The online deep learning forecasting model, the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the online deep learning forecasting model are the same as the offline deep learning forecasting model, and the offline deep learning forecasting model is the same as the offline deep learning forecasting model.
- the weight parameters and bias parameters of each layer of the model are used as the initial value of the weight parameter and the initial value of the bias parameter of the corresponding layer of the online deep learning prediction model, and an online training algorithm is adopted.
- the error between the outputs of the deep learning prediction model, the weight parameter and the bias parameter of the last layer of the online deep learning prediction model are corrected online; the deep learning correction model modeling module adopts LSTM to establish the deep learning correction model , the input of a single neuron, the number of neurons, the number of unit nodes and the number of network layers of the deep learning correction model are the same as the online deep learning prediction model.
- the error between the outputs of the learning correction model is corrected in real time, and the weight parameters and bias parameters of each layer of the deep learning correction model are corrected in real time; when the self-correction module meets the preset conditions, the The weight parameters and bias parameters of the layers replace the weight parameters and bias parameters of the corresponding layers of the online deep learning prediction model; wherein, the historical data input by the deep learning correction model is larger than that input by the online deep learning prediction model. of historical data.
- the online correction of the weight parameters and bias parameters of the last layer of the online deep learning prediction model specifically, the online correction of part of the weights of the last layer of the online deep learning prediction model parameters and partial bias parameters.
- the complex industrial system is an alumina production system
- the online deep learning prediction model is used to perform real-time prediction on the detection error of the caustic alkali concentration of the alumina production system; the caustic alkali
- the concentration detection error is the difference between the assay value of caustic alkali concentration and the measured value of the caustic alkali concentration online detection instrument.
- Each module in the above-mentioned intelligent forecasting device for complex industrial systems can be implemented in whole or in part by software, hardware and combinations thereof.
- the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
- an intelligent forecasting device for complex industrial systems based on adaptive deep learning for implementing the intelligent forecasting methods in the above embodiments including: terminal-side sub-devices, edge side sub-device and cloud-side sub-device; the terminal-side sub-device is used to collect the input data and output data of the complex industrial system; the edge-side sub-device uses the online deep learning prediction model to analyze the complex industrial system The cloud-side sub-device is used to train the deep learning correction model, and realize the correction of the online deep learning forecasting model by the deep learning correction model.
- a computer-readable storage medium which stores a computer program, and when the program is executed by a processor, implements the intelligent forecasting method for a complex industrial system in each of the foregoing embodiments.
- the intelligent forecasting method, device and equipment for complex industrial systems proposed in the embodiments of the present invention aim at the problems of low forecasting accuracy and poor real-time forecasting of complex industrial systems, and establish an offline deep learning forecasting model and an online deep learning forecasting model. , deep learning correction model and self-correction mechanism to achieve accurate real-time forecasting of complex industrial systems.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Feedback Control In General (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
一种基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质。复杂工业系统智能预报方法包括:建立复杂工业系统的动态模型(S1);利用所述动态模型建立离线深度学习预报模型(S2);利用所述离线深度学习预报模型建立在线深度学习预报模型(S3);采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型(S4);利用所述深度学习校正模型校正所述在线深度学习预报模型(S5);其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时预报。针对复杂工业系统预报精度较低及预报实时性差的问题,建立了包括离线深度学习预报模型、在线深度学习预报模型、深度学习校正模型及自校正的机制,实现了复杂工业系统的精确实时预报。
Description
本发明属于工业人工智能技术领域,涉及基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质。
为了实现生产过程的闭环优化决策,需要对表征产品质量、效率、消耗的生产指标和关键工艺参数在线预报。由于工业过程的闭环优化决策的时间周期短,要求生产指标和关键工艺参数的预报模型在决策时间周期给出预报值。这就要求深度学习预报模型的训练数据集不能过大,训练算法不能耗时过长。加上生产制造过程的复杂性,造成生产指标和关键工艺参数与相关的生产过程输入与输出变量是复杂的动态系统。该动态系统往往具有强非线性,多变量强耦合,模型结构与输入与输出变量的阶次未知,甚至变化,生产原料等生产边界条件的变化、生产过程中的物质流、信息流、能源流的相互作用,使该动态系统的特性随生产时间而发生未知变化,导致该系统的输入、输出数据处于变化的、开放的、不确定的信息空间,使得已有的完备信息空间的深度学习技术无法应用于上述复杂的工业动态系统中建立该系统的预报模型。
发明内容
本发明旨在至少在一定程度上解决相关技术中的技术问题之一。本发明的技术方案如下:
一种基于自适应深度学习的复杂工业系统智能预报方法,包括如下步骤:
建立复杂工业系统的动态模型;
利用所述动态模型建立离线深度学习预报模型;
利用所述离线深度学习预报模型建立在线深度学习预报模型;
采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型;
利用所述深度学习校正模型校正所述在线深度学习预报模型;
其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时 预报。
进一步,作为优选,所述建立复杂工业系统的动态模型包括:确定所述动态模型的输入变量和输出变量,所述输出变量为被预报的变量;所述利用所述动态模型建立离线深度学习预报模型包括:采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数;所述利用所述离线深度学习预报模型建立在线深度学习预报模型包括:采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用在线训练算法,根据所述标签数据与所述在线深度学习预报模型输出之间的误差,在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数;所述采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型包括:采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,通过训练算法,根据所述标签数据与所述深度学习校正模型输出之间的误差,实时校正所述深度学习校正模型的各层的权值参数和偏置参数;所述利用所述深度学习校正模型校正所述在线深度学习预报模型包括:当满足预设条件时,采用所述深度学习校正模型的各层的权值参数和偏置参数替换所述在线深度学习预报模型的相应层的权值参数和偏置参数;其中,所述深度学习校正模型输入的历史数据比所述在线深度学习预报模型输入的历史数据多。
进一步,作为优选,所述在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数,具体为在线校正所述在线深度学习预报模型的最后一层的部分权值参数和部分偏置参数。
进一步,作为优选,所述复杂工业系统为氧化铝制取系统,所述在线深度学习预报模型用于对所述氧化铝制取系统的苛性碱浓度检测误差进行实时预报;所述苛性碱浓度检测误差为苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量 值之差。
一种基于自适应深度学习的复杂工业系统智能预报装置,包括:
动态模型建模模块,用于建立复杂工业系统的动态模型;
离线深度学习预报模型建模模块,用于利用所述动态模型建立离线深度学习预报模型;
在线深度学习预报模型建模模块,用于利用所述离线深度学习预报模型建立在线深度学习预报模型;
深度学习校正模型建模模块,用于采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型;
自校正模块,用于利用所述深度学习校正模型校正所述在线深度学习预报模型;
其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时预报。
进一步,作为优选,所述动态模型建模模块确定所述动态模型的输入变量和输出变量,所述输出变量为被预报的变量;所述离线深度学习预报模型建模模块采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数;所述在线深度学习预报模型建模模块采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用在线训练算法,根据所述标签数据与所述在线深度学习预报模型输出之间的误差,在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数;所述深度学习校正模型建模模块采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,通过训练算法,根据所述标签数据与所述深度学习校正模型输出之间的误差,实时校正所述深度学习校正模 型的各层的权值参数和偏置参数;所述自校正模块在满足预设条件时,采用所述深度学习校正模型的各层的权值参数和偏置参数替换所述在线深度学习预报模型的相应层的权值参数和偏置参数;其中,所述深度学习校正模型输入的历史数据比所述在线深度学习预报模型输入的历史数据多。
进一步,作为优选,所述在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数,具体为在线校正所述在线深度学习预报模型的最后一层的部分权值参数和部分偏置参数。
进一步,作为优选,所述复杂工业系统为氧化铝制取系统,所述在线深度学习预报模型用于对所述氧化铝制取系统的苛性碱浓度检测误差进行实时预报;所述苛性碱浓度检测误差为苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量值之差。
一种用于实现上述智能预报方法的基于自适应深度学习的复杂工业系统智能预报设备,所述设备包括:端侧子设备、边缘侧子设备和云侧子设备;
所述端侧子设备用于采集所述复杂工业系统的输入数据和输出数据;
所述边缘侧子设备利用所述在线深度学习预报模型对所述复杂工业系统的参数进行实时预报;
所述云侧子设备用于训练所述深度学习校正模型,并实现所述深度学习校正模型对所述在线深度学习预报模型的校正。
一种计算机可读存储介质,其存储有计算机程序,所述程序被处理器执行时实现上述复杂工业系统智能预报方法。
本发明针对复杂工业系统预报精度较低及预报实时性差的问题,建立了包括离线深度学习预报模型、在线深度学习预报模型、深度学习校正模型及自校正的机制,实现了复杂工业系统的精确实时预报。
图1为本发明实施例的复杂工业系统智能预报方法实现流程图;
图2为本发明一个实施例的苛性碱浓度检测误差智能预报方法实现流程图;
图3为输入数据时间序列窗口长度取不同值时在线深度学习预报模型的预报误差;
图4为本发明一个实施例的复杂工业系统智能预报装置的结构示意图;
图5为本发明一个实施例的复杂工业系统智能预报设备的结构示意图。
下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
图1为本发明实施例的复杂工业系统智能预报方法实现流程图,该方法包括以下步骤:
S1:建立复杂工业系统的动态模型。
具体的,通过工业过程的机理分析,以需要预报的生产指标或关键工艺参数为工业系统动态模型的输出变量,影响该输出变量的工业过程的输入和输出为该动态模型的输入变量,并将该动态模型的输出历史数据以及预报误差历史数据作为该动态模型的输入变量,采用未知常数n表示该动态系统的输入与输出变量的未知变化的阶次。采用下式表示该工业系统的动态模型:
s(k)=f(s(k-1),…,s(k-n),y
i(k),…,y
i(k-n+1),u
i(k),…,u
i(k-n+1),Δs(k-1),…,Δs(k-n)) (1)
其中,f是未知变化的非线性函数;s(k)为该动态模型k时刻的输出;y
i(k)为工业过程k时刻的第i个输出,u
i(k)为工业过程k时刻的第i个输入,i=1,…,m;
为k-1时刻的预报误差,即k-1时刻该动态模型的输出s(k-1)与预报模型的输出
之差。
S2:利用所述动态模型建立离线深度学习预报模型。
具体的,采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数。
步骤S2包括步骤S21和S22。
步骤S21为:采用LSTM结构建立离线深度学习预报模型,将LSTM的初始网络层数设置为1,通过训练算法,根据标签数据与离线深度学习预报模型输出之差确定所述LSTM的神经元个数和单元节点数。
具体的,选择(1)式中的输入变量作为第j个单个神经元的输入x
j(k+j-n)(j=1,…,n),阶次n为神经元的个数,即:
x
j(k+j-n)=[s(k+j-n-1),y
i(k+j-n),u
i(k+j-n),Δs(k+j-n-1))]
T (2)
其中,j=1,…,n;i=1,…,m。
以工业系统动态模型(1)的输出数据s(k)作为标签数据,采用(1)式的输入、输出数据组成大数据样本,采用离线训练算法,使标签数据与离线深度学习预报模型输出之差尽可能小,确定LSTM的神经元个数n、单元节点数h。
步骤S22为:固定所述LSTM的神经元个数和单元节点数,改变所述LSTM的网络层数,根据不同网络层数对应的所述标签数据与所述离线深度学习预报模型输出之差,选择LSTM的网络层数。
具体的,固定所述LSTM的单个神经元个数n和LSTM单元节点数h,通过增加LSTM的网络层数来减小离线深度学习预报模型的输出与标签数据的误差,使该误差尽可能小,确定LSTM的网络层数以及各层权值参数和偏置参数。
S3:利用所述离线深度学习预报模型建立在线深度学习预报模型。
具体的,采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用固定数据量的时间序列N来在线校正该在线深度学习预报模型最后一层的权值参数和偏置参数,来保证在线深度学习预报模型在确定的优化决策时间周期内完成预报算法。采用训练算法,通过使预报误差尽可能小,确定N。采用时间序列长度为N的数据集,采用递推算法,即k时刻在线深度学习预报模型的输入数据的时刻序列为(k-N+1),…,k;(k+1)时刻在线深度学习预报模型的输入数据的时刻序列为(k-N+2),…,(k+1)。在线深度学习预报模型用于对复杂工业系统的参数进行实时预报。
S4:采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型。
具体的,采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的 单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,采用当前时刻以及以前所有时刻的模型(1)式的输入数据作为深度学习校正模型的输入数据,训练深度学习校正模型的各层的所有权值参数和偏置参数,获得深度学习校正模型的预报值
和预报误差
S5:利用所述深度学习校正模型校正所述在线深度学习预报模型。
具体的,利用所述深度学习校正模型自适应校正所述在线深度学习预报模型,设定预报误差的区间上界为δ,当在线深度学习预报模型的预报误差|Δs(k)|≥δ,且深度学习校正模型的预报误差
采用深度学习校正模型的各层的权值参数和偏置参数替换在线深度学习预报模型的相应层的权值参数和偏置参数,保证在线深度学习预报模型的预报误差在设定预报误差的区间内,即|Δs(k)|<δ。
进一步的,在一个实施例中,复杂工业系统智能预报方法可用于氧化铝制取系统的苛性碱浓度检测误差的预报。
氧化铝具有硬度高、熔点高等优良特性,常用于冶炼金属铝和制造耐火材料,是军工、航天和国民经济中具有支撑作用的战略资源。目前,制取氧化铝的主要方法是拜耳法,其基本工艺流程通常是将碎铝土矿按配比要求加入石灰和苛性碱溶液进行磨制,再利用苛性碱溶液在一定温度、一定压力条件下溶出铝土矿制得铝酸钠溶液,铝酸钠溶液净化后经过降温、添加晶种、搅拌分解析出氢氧化铝晶体,析出的氢氧化铝经分离、洗涤、焙烧后得到氧化铝。分离氢氧化铝后的母液(主要成分是苛性碱)经蒸发过程后再重新溶出新的铝土矿,进入下一循环。
氧化铝溶液的苛性碱浓度是氧化铝蒸发过程的关键工艺指标,关系到氧化铝的最终产品质量。日常苛性碱浓度检测是依靠人工按固定周期取样后再化验的手段获得准确的苛性碱浓度值,然而由于取样间隔和化验所需时间较长,苛性碱浓度的检测存在严重滞后性,无法实现蒸发过程的运行优化控制。
为了实现蒸发过程的运行优化控制,一些氧化铝企业以昂贵的价格引进苛性碱浓度在线测量装置。实际生产中,铝土矿品位的变化和生产过程运行的变化导致该检测装置测得的苛性碱浓度与化验结果差距大,无法使用。由于该误差的动态特性具有未知强非线性,模型阶次未知,生产原料等生产边界条件时常波动、各工艺流程和物料间相互作用,因此该误差动态系统的特性随生产时间而发生未 知变化,导致该系统的输入、输出数据处于变化的、开放的、不确定的信息空间,使得已有的完备信息空间的深度学习技术难以应用于氧化铝蒸发过程苛性碱浓度误差预报动态系统中。除此之外,由于氧化铝制取工业过程的运行优化决策的时间周期短,要求生产指标和关键工艺参数的预报模型在决策时间周期给出预报值。这就要求深度学习预报模型的训练数据集不能过大,训练算法不能耗时过长。
本发明实施例针对氧化铝制取系统预报精度较低及预报实时性差的问题,建立了包括离线深度学习预报模型、在线深度学习预报模型、深度学习校正模型及自校正的机制,实现了氧化铝制取系统的精确实时预报。
图2为本发明一个实施例的苛性碱浓度检测误差智能预报方法实现流程图,该方法包括以下步骤:
S1’:建立苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量值的检测误差动态模型。
具体的,由于苛性碱浓度在线检测仪表以氧化铝溶液折光度和温度为输入,通过模型产生浓度测量值,因此该检测误差动态模型的输入包括氧化铝溶液的折光度和温度,并将苛性碱浓度化验值和苛性碱浓度检测仪表测量值之差的历史值作为该检测误差动态模型的输入,采用未知常数n表示该动态系统的输入与输出变量的未知阶次,建立苛性碱浓度检测误差动态模型如下:
S2’:利用所述检测误差动态模型建立离线深度学习预报模型。
具体的,采用LSTM建立所述离线深度学习预报模型,将所述检测误差动态模型的所述输入变量作为LSTM的输入,将所述检测误差动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数。
步骤S2’包括步骤S21’和S22’。
步骤S21’为:采用LSTM结构建立离线深度学习预报模型,将LSTM的初始网络层数设置为1,通过训练算法,根据标签数据与离线深度学习预报模型输出之差确定所述LSTM的神经元个数和单元节点数。
具体的,选择(3)式中第j个输入变量作为第j个单个神经元的输入x
j(k+j-n)(j=1,…,n),即:
x
j(k+j-n)=[y
1(k+j-n),y
2(k+j-n),Δr(k+j-n-1)]
T (4)
其中,j=1,…,n;n为神经元的个数。
选择层数为1的LSTM神经网络,利用(3)式所示输入输出变量所构成的大数据样本,以苛性碱浓度化验值与检测仪表测量值的误差值Δr(k)作为标签数据,采用训练算法,使k时刻的标签数据Δr(k)与k时刻的离线深度学习预报模型输出
的误差尽可能小来确定n和
训练算法的目标函数为:
其中,M表示训练数据的数据量。
标签数据Δr(k):
h(k)=o
k*tanh(C(k)) (8)
o
k=σ(W
o·[h(k-1),x
j(k)]
T+b
o) (9)
其中h(k-1)是第(n-1)个神经元的输出,[h(k-1),x
j(k)]
T为
向量,W
o和b
o为神经网络第一层的连接权和偏置,W
o为
矩阵,b
o为
向量。σ为sigmoid函数,σ(z)=(1+e
-z)
-1,z为向量[W
o·[h(k-1),x
j(k)]
T+b
o]的元素。
f
k=σ(W
f·[h(k-1),x
j(k)]
T+b
f)
i
k=σ(W
i·[h(k-1),x
j(k)]
T+b
i) (11)
取n=1,2…22,
使用(5)—(11)式采用梯度下降算法使(5)式极小,实验过程中当神经元的个数n为20且LSTM的单元节点数
为180时测试误差最小,因此确定神经元的个数n为20,LSTM的单元节点数
为180。
其中,h
L(k-1)为第L层LSTM神经网络的第19个神经元的输出,h
L-1(k)为第L-1层LSTM神经网络的第20个神经元的输出,也是第L层LSTM的第20个神经元的输入。
C
L(k)为长期记忆状态,C
L(k)为180×1向量
取L=1,2,3,4,使用(5)、(6)和(12)—(16)式采用梯度下降算法使(5)式极小,实验结果如表1所示,当LSTM神经网络的层数L为2时预报精度最高且训练时间较短,因此确定神经网络层数为两层,同时确定离线深度学习预报模型的各层连接权参数和偏置参数。
表1 测试误差与神经网络中LSTM单元层数
S3’:利用所述离线深度学习预报模型建立在线深度学习预报模型。
具体的,采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线 深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,并在线校正在线深度学习预报模型的第二层的连接权
和偏置
在线深度学习预报模型如下:
式中,
为
的k时刻的校正值,
为
的k时刻的校正值,h
2(k)为第2层LSTM单元最后一个神经元输出。为了确保在规定的预报周期内在线完成预报算法,通过遍历确定在线深度学习预报模型的输入数据时间序列窗口长度N。
目标函数为:
取N=500,…,900,采用上述算法校正
和
由(17)和(18)式计算预报误差实验结果如图3所示。当N小于820时无法满足预报模型的精度需求,当N大于820时会产生冗余、增加网络计算量,当N为820时预报误差最小,因此确定在线深度学习预报模型输入数据序列的时间长度N=820。
相应的,(k+1)时刻的苛性碱浓度检测误差的在线深度学习预报模型为:
(k+1)时刻在线深度学习预报模型利用N为820的时间序列(k-818),(k-817),…,(k+1)时刻的输入数据,采用如下算法校正权值参数
和偏置参数
由(21)式求取(k+1)时刻苛性碱浓度检测误差预 报值
其中η表示在线深度学习预报模型中参数校正的学习率,η=0.0005。
S4’:采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型。
具体的,采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同。采用当前k时刻以及以前所有时刻即k,…,2,1的模型(3)式的所有输入数据作为深度学习校正模型的输入数据,采用下列目标函数和训练算法校正该深度学习校正模型的第一层和第二层的所有权值参数和偏置参数。
目标函数为:
其中,η表示校正模型中参数校正的学习率,η=0.0005。采用上述校正算法可校正该深度学习校正模型的第一层和第二层其余连接权参数 W
f,W
i,W
C,W
o,W
d,
和偏置参数b
f,b
i,b
C,b
o,b
d,
S5’:利用所述深度学习校正模型校正在线深度学习预报模型。
具体的,利用所述深度学习校正模型自适应校正所述在线深度学习预报模型,设定预报误差的区间上界为δ,δ=1.5g/l,采样时刻i的区间为[k,k-1,…k-99]。连续100个最新时刻的采样点内,当在线深度学习预报模型中预报误差
未超过区间上界的采样点个数小于99而深度学习校正模型中存在99个采样点的预报误差
未超区间上界时,采用深度学习校正模型的各层权值参数和偏置参数校正在线深度学习预报模型的各层权值参数和偏置参数,保证在线深度学习预报模型的预报误差在设定预报误差的区间范围内。
将本发明实施例的苛性碱浓度检测误差预报方法应用在山西某氧化铝厂蒸发过程中的效果如表2所示。
表2中的仪表测量值为苛性碱浓度仪表在线测量值,补偿后的仪表值为苛性碱浓度仪表在线测量值与苛性碱浓度检测误差在线深度学习预报模型输出的预报值之和。表2分别统计了仪表测量值、补偿后的仪表值与苛性碱浓度化验值的均方根误差RMSE以及在生产工艺规定的误差区间范围内的合格率。由表2可以看出,采用本发明实施例的苛性碱浓度检测误差预报方法对仪表测量值进行补偿后,可将苛性碱浓度仪表测量值与化验值之间的RMSE由11.25下降为0.50,合格率由10.75%提升为99.62%,为实现氧化铝蒸发过程的闭环运行优化控制创造了条件。
表2 苛性碱浓度在线深度学习预报模型应用效果
在一个实施例中,如图4所示,提供了一种基于自适应深度学习的复杂工业系统智能预报装置,包括:动态模型建模模块、离线深度学习预报模型建模模块、在线深度学习预报模型建模模块、深度学习校正模型建模模块和自校正模块,其 中:
动态模型建模模块用于建立复杂工业系统的动态模型;
离线深度学习预报模型建模模块用于利用所述动态模型建立离线深度学习预报模型;
在线深度学习预报模型建模模块用于利用所述离线深度学习预报模型建立在线深度学习预报模型;
深度学习校正模型建模模块用于采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型;
自校正模块用于利用所述深度学习校正模型校正所述在线深度学习预报模型;
其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时预报。
在其中一个实施例中,所述动态模型建模模块确定所述动态模型的输入变量和输出变量,所述输出变量为被预报的变量;所述离线深度学习预报模型建模模块采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数;所述在线深度学习预报模型建模模块采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用在线训练算法,根据所述标签数据与所述在线深度学习预报模型输出之间的误差,在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数;所述深度学习校正模型建模模块采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,通过训练算法,根据所述标签数据与所述深度学习校正模型输出之间的误差,实时校正所述深度学习校正模型的各层的权值参数和偏置参数;所述自校正模块在满足预设条件时,采用所 述深度学习校正模型的各层的权值参数和偏置参数替换所述在线深度学习预报模型的相应层的权值参数和偏置参数;其中,所述深度学习校正模型输入的历史数据比所述在线深度学习预报模型输入的历史数据多。
在其中一个实施例中,所述在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数,具体为在线校正所述在线深度学习预报模型的最后一层的部分权值参数和部分偏置参数。
在其中一个实施例中,所述复杂工业系统为氧化铝制取系统,所述在线深度学习预报模型用于对所述氧化铝制取系统的苛性碱浓度检测误差进行实时预报;所述苛性碱浓度检测误差为苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量值之差。
关于复杂工业系统智能预报装置的具体限定可以参见上文中对于复杂工业系统智能预报方法的限定,在此不再赘述。上述复杂工业系统智能预报装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中,如图5所示,提供了一种用于实现上述各实施例中的智能预报方法的基于自适应深度学习的复杂工业系统智能预报设备,包括:端侧子设备、边缘侧子设备和云侧子设备;所述端侧子设备用于采集所述复杂工业系统的输入数据和输出数据;所述边缘侧子设备利用所述在线深度学习预报模型对所述复杂工业系统的参数进行实时预报;所述云侧子设备用于训练所述深度学习校正模型,并实现所述深度学习校正模型对所述在线深度学习预报模型的校正。
在一个实施例中,提供了一种计算机可读存储介质,其存储有计算机程序,所述程序被处理器执行时实现上述各实施例中的复杂工业系统智能预报方法。
在不相互矛盾的情况下,本领域的技术人员可以将本说明书中描述的不同实施例以及不同实施例的特征进行结合和组合。
综上所述,本发明实施例提出的复杂工业系统智能预报方法、装置和设备针对复杂工业系统预报精度较低及预报实时性差的问题,建立了包括离线深度学习预报模型、在线深度学习预报模型、深度学习校正模型及自校正的机制,实现了复杂工业系统的精确实时预报。
以上所述,仅为本发明的具体实施方式,但本发明的保护范围并不局限于此,任何熟悉本技术领域的技术人员在本发明揭露的技术范围内,可轻易想到其各种变化或替换,这些都应涵盖在本发明的保护范围之内。因此,本发明的保护范围应以所述权利要求的保护范围为准。
Claims (10)
- 一种基于自适应深度学习的复杂工业系统智能预报方法,其特征在于,所述方法包括:建立复杂工业系统的动态模型;利用所述动态模型建立离线深度学习预报模型;利用所述离线深度学习预报模型建立在线深度学习预报模型;采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型;利用所述深度学习校正模型校正所述在线深度学习预报模型;其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时预报。
- 根据权利要求1所述的方法,其特征在于,所述建立复杂工业系统的动态模型,包括:确定所述动态模型的输入变量和输出变量,所述输出变量为被预报的变量;所述利用所述动态模型建立离线深度学习预报模型,包括:采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数;所述利用所述离线深度学习预报模型建立在线深度学习预报模型,包括:采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用在线训练算法,根据所述标签数据与所述在线深度学习预报模型输出之间的误差,在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数;所述采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型,包括:采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,通过训练算法,根据所述标签数据与所述深度学习校正模型输出之间的误差,实时校正所述深度学习校正模型的各层的权值参数和偏置参数;所述利用所述深度学习校正模型校正所述在线深度学习预报模型,包括:当满足预设条件时,采用所述深度学习校正模型的各层的权值参数和偏置参数替换所述在线深度学习预报模型的相应层的权值参数和偏置参数;其中,所述深度学习校正模型输入的历史数据比所述在线深度学习预报模型输入的历史数据多。
- 根据权利要求2所述的方法,其特征在于,所述在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数,具体为在线校正所述在线深度学习预报模型的最后一层的部分权值参数和部分偏置参数。
- 根据权利要求1-3任一所述的方法,其特征在于,所述复杂工业系统为氧化铝制取系统,所述在线深度学习预报模型用于对所述氧化铝制取系统的苛性碱浓度检测误差进行实时预报;所述苛性碱浓度检测误差为苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量值之差。
- 一种基于自适应深度学习的复杂工业系统智能预报装置,其特征在于,所述装置包括:动态模型建模模块,用于建立复杂工业系统的动态模型;离线深度学习预报模型建模模块,用于利用所述动态模型建立离线深度学习预报模型;在线深度学习预报模型建模模块,用于利用所述离线深度学习预报模型建立在线深度学习预报模型;深度学习校正模型建模模块,用于采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型;自校正模块,用于利用所述深度学习校正模型校正所述在线深度学习预报模型;其中,所述在线深度学习预报模型用于对所述复杂工业系统的参数进行实时预报。
- 根据权利要求5所述的装置,其特征在于,所述动态模型建模模块确定所述动态模型的输入变量和输出变量,所述输出变量为被预报的变量;所述离线深度学习预报模型建模模块采用LSTM建立所述离线深度学习预报模型,将所述动态模型的所述输入变量作为LSTM的输入,将所述动态模型的输 出数据作为标签数据,采用离线训练算法,根据所述标签数据与所述离线深度学习预报模型输出之间的误差,确定LSTM的神经元个数、单元节点数、网络层数和各层的权值参数及偏置参数;所述在线深度学习预报模型建模模块采用LSTM建立所述在线深度学习预报模型,所述在线深度学习预报模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述离线深度学习预报模型相同,将所述离线深度学习预报模型的各层的权值参数和偏置参数作为所述在线深度学习预报模型的相应层的权值参数初始值和偏置参数初始值,采用在线训练算法,根据所述标签数据与所述在线深度学习预报模型输出之间的误差,在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数;所述深度学习校正模型建模模块采用LSTM建立所述深度学习校正模型,所述深度学习校正模型的单个神经元的输入、神经元个数、单元节点数和网络层数均与所述在线深度学习预报模型相同,通过训练算法,根据所述标签数据与所述深度学习校正模型输出之间的误差,实时校正所述深度学习校正模型的各层的权值参数和偏置参数;所述自校正模块在满足预设条件时,采用所述深度学习校正模型的各层的权值参数和偏置参数替换所述在线深度学习预报模型的相应层的权值参数和偏置参数;其中,所述深度学习校正模型输入的历史数据比所述在线深度学习预报模型输入的历史数据多。
- 根据权利要求6所述的装置,其特征在于,所述在线校正所述在线深度学习预报模型的最后一层的权值参数和偏置参数,具体为在线校正所述在线深度学习预报模型的最后一层的部分权值参数和部分偏置参数。
- 根据权利要求5-7任一所述的装置,其特征在于,所述复杂工业系统为氧化铝制取系统,所述在线深度学习预报模型用于对所述氧化铝制取系统的苛性碱浓度检测误差进行实时预报;所述苛性碱浓度检测误差为苛性碱浓度化验值与苛性碱浓度在线检测仪表的测量值之差。
- 一种用于实现权利要求1-4所述方法的基于自适应深度学习的复杂工业系统智能预报设备,其特征在于,所述设备包括:端侧子设备、边缘侧子设备和云 侧子设备;所述端侧子设备用于采集所述复杂工业系统的输入数据和输出数据;所述边缘侧子设备利用所述在线深度学习预报模型对所述复杂工业系统的参数进行实时预报;所述云侧子设备用于训练所述深度学习校正模型,并实现所述深度学习校正模型对所述在线深度学习预报模型的校正。
- 一种计算机可读存储介质,其存储有计算机程序,其特征在于,所述程序被处理器执行时实现如权利要求1-4中任一所述的方法。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US18/208,881 US20230325689A1 (en) | 2020-12-10 | 2023-06-12 | Adaptive deep learning-based intelligent prediction method, apparatus, and device for complex industrial system, and storage medium |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011435304.2 | 2020-12-10 | ||
CN202011435304.2A CN112560338B (zh) | 2020-12-10 | 2020-12-10 | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US18/208,881 Continuation US20230325689A1 (en) | 2020-12-10 | 2023-06-12 | Adaptive deep learning-based intelligent prediction method, apparatus, and device for complex industrial system, and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022121932A1 true WO2022121932A1 (zh) | 2022-06-16 |
Family
ID=75060615
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/136373 WO2022121932A1 (zh) | 2020-12-10 | 2021-12-08 | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 |
Country Status (3)
Country | Link |
---|---|
US (1) | US20230325689A1 (zh) |
CN (1) | CN112560338B (zh) |
WO (1) | WO2022121932A1 (zh) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115323440A (zh) * | 2022-09-30 | 2022-11-11 | 湖南力得尔智能科技股份有限公司 | 基于ai神经网络深度自学习的铝电解全息化闭环控制系统 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE102020114339A1 (de) * | 2020-05-28 | 2021-12-02 | Ebm-Papst Mulfingen Gmbh & Co. Kg | Verfahren zum Betreiben eines Ventilatorsystems und Ventilatorsystem mit einem rückwärtsgekrümmten Radialventilator |
CN112560338B (zh) * | 2020-12-10 | 2022-03-25 | 东北大学 | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 |
Citations (7)
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 |
CN106934209A (zh) * | 2017-01-19 | 2017-07-07 | 上海电气电站环保工程有限公司 | 一种燃煤电站烟气含氧量在线预测方法 |
CN109033450A (zh) * | 2018-08-22 | 2018-12-18 | 太原理工大学 | 基于深度学习的电梯设备故障预测方法 |
CN109272146A (zh) * | 2018-08-23 | 2019-01-25 | 河海大学 | 一种基于深度学习模型和bp神经网络校正的洪水预测方法 |
US20190102693A1 (en) * | 2017-09-29 | 2019-04-04 | Facebook, Inc. | Optimizing parameters for machine learning models |
CN109784471A (zh) * | 2018-10-24 | 2019-05-21 | 浙江华云信息科技有限公司 | 一种基于多预测模型的95598话务工单预测与异动预警方法 |
CN112560338A (zh) * | 2020-12-10 | 2021-03-26 | 东北大学 | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 |
Family Cites Families (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1546982A (zh) * | 2003-12-08 | 2004-11-17 | 中国铝业股份有限公司 | 一种氧化铝生产过程中苛性比值与溶出率在线检测方法 |
CN101339408A (zh) * | 2008-08-15 | 2009-01-07 | 中南大学 | 铝酸钠溶液连续碳酸化分解过程智能控制系统 |
CN102629104B (zh) * | 2011-12-01 | 2013-10-02 | 燕山大学 | 一种水泥回转窑煅烧预测控制系统及方法 |
CN106055844B (zh) * | 2016-07-06 | 2019-02-12 | 中南大学 | 基于bp神经网络的镍基高温合金微观组织预测控制方法 |
CN109325417B (zh) * | 2018-08-23 | 2021-09-07 | 东北大学 | 一种基于深度神经网络的工业过程故障工况诊断方法 |
CN109299812B (zh) * | 2018-08-23 | 2021-09-24 | 河海大学 | 一种基于深度学习模型和knn实时校正的洪水预测方法 |
CN111652402B (zh) * | 2019-03-04 | 2023-04-14 | 湖南师范大学 | 一种光纤预制棒沉积工序大数据智能优化方法 |
CN110598737B (zh) * | 2019-08-06 | 2023-02-24 | 深圳大学 | 一种深度学习模型的在线学习方法、装置、设备及介质 |
CN110502840B (zh) * | 2019-08-26 | 2021-05-28 | 南京航空航天大学 | 航空发动机气路参数在线预测方法 |
CN110866659B (zh) * | 2019-12-09 | 2023-08-15 | 东北大学 | 一种拜耳法氧化铝生产高压管道溶出过程中苛性比值在线预测方法 |
CN111047094A (zh) * | 2019-12-12 | 2020-04-21 | 国网浙江省电力有限公司 | 一种基于深度学习算法的抄表数据异常分析方法 |
CN111260124A (zh) * | 2020-01-11 | 2020-06-09 | 大连理工大学 | 一种基于注意力机制深度学习的混沌时间序列预测方法 |
CN111582562A (zh) * | 2020-04-20 | 2020-08-25 | 杭州意能电力技术有限公司 | 一种基于优化控制平台的神经网络预测控制方法 |
CN111768000A (zh) * | 2020-06-23 | 2020-10-13 | 中南大学 | 在线自适应微调深度学习的工业过程数据建模方法 |
CN111913803B (zh) * | 2020-07-21 | 2023-12-29 | 哈尔滨工程大学 | 一种基于akx混合模型的服务负载细粒度预测方法 |
-
2020
- 2020-12-10 CN CN202011435304.2A patent/CN112560338B/zh active Active
-
2021
- 2021-12-08 WO PCT/CN2021/136373 patent/WO2022121932A1/zh active Application Filing
-
2023
- 2023-06-12 US US18/208,881 patent/US20230325689A1/en active Pending
Patent Citations (7)
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 |
CN106934209A (zh) * | 2017-01-19 | 2017-07-07 | 上海电气电站环保工程有限公司 | 一种燃煤电站烟气含氧量在线预测方法 |
US20190102693A1 (en) * | 2017-09-29 | 2019-04-04 | Facebook, Inc. | Optimizing parameters for machine learning models |
CN109033450A (zh) * | 2018-08-22 | 2018-12-18 | 太原理工大学 | 基于深度学习的电梯设备故障预测方法 |
CN109272146A (zh) * | 2018-08-23 | 2019-01-25 | 河海大学 | 一种基于深度学习模型和bp神经网络校正的洪水预测方法 |
CN109784471A (zh) * | 2018-10-24 | 2019-05-21 | 浙江华云信息科技有限公司 | 一种基于多预测模型的95598话务工单预测与异动预警方法 |
CN112560338A (zh) * | 2020-12-10 | 2021-03-26 | 东北大学 | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115323440A (zh) * | 2022-09-30 | 2022-11-11 | 湖南力得尔智能科技股份有限公司 | 基于ai神经网络深度自学习的铝电解全息化闭环控制系统 |
CN115323440B (zh) * | 2022-09-30 | 2023-04-07 | 湖南力得尔智能科技股份有限公司 | 基于ai神经网络深度自学习的铝电解全息化闭环控制系统 |
Also Published As
Publication number | Publication date |
---|---|
CN112560338B (zh) | 2022-03-25 |
CN112560338A (zh) | 2021-03-26 |
US20230325689A1 (en) | 2023-10-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022121932A1 (zh) | 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质 | |
WO2022121956A1 (zh) | 基于深度学习的复杂工业系统预报模型构建方法、装置、设备及存储介质 | |
WO2022121944A1 (zh) | 工业过程运行指标智能预报方法、装置、设备及存储介质 | |
WO2022121923A1 (zh) | 复杂工业过程数字孪生系统智能建模方法、装置、设备及存储介质 | |
CN108932567B (zh) | 一种基于卷积神经网络的水泥烧成过程多能耗指标预测方法 | |
CN109839825B (zh) | 一种稀土萃取过程组分含量的预测控制方法及系统 | |
CN101863088B (zh) | 一种橡胶混炼过程中门尼粘度的预报方法 | |
CN103389472B (zh) | 一种基于nd-ar模型的锂离子电池循环寿命的预测方法 | |
WO2020147461A1 (zh) | 一种基于数据驱动的溶剂脱沥青装置建模及优化方法 | |
WO2021114320A1 (zh) | 一种oica和rnn融合模型的污水处理过程故障监测方法 | |
CN111680786A (zh) | 一种基于改进权重门控单元的时序预测方法 | |
CN116088349B (zh) | 基于lstm的火箭半实物仿真数据异常评估方法及决策装置 | |
CN110222825B (zh) | 一种水泥成品比表面积预测方法及系统 | |
CN111860839A (zh) | 基于多信号融合及Adam优化算法的岸桥故障监测方法 | |
CN108614533A (zh) | 一种基于narx模型和时滞估计的神经网络建模方法 | |
CN111222798B (zh) | 一种复杂工业过程关键性指标软测量方法 | |
CN103279030B (zh) | 基于贝叶斯框架的动态软测量建模方法及装置 | |
CN110763830B (zh) | 一种水泥熟料游离氧化钙含量预测方法 | |
US11987855B2 (en) | Method and system for determining converter tapping quantity | |
Zhou et al. | Anode effect prediction of aluminum electrolysis using GRNN | |
CN112257893A (zh) | 一种考虑监测误差的复杂机电系统健康状态预测方法 | |
CN112364527B (zh) | 一种基于aliesn在线学习算法的脱丁烷塔软测量建模方法 | |
CN109725526A (zh) | 一种多变量半自适应预测控制方法 | |
CN107977742B (zh) | 一种中长期电力负荷预测模型的构建方法 | |
CN113609772B (zh) | 基于卷积准循环神经网络的水泥成品比表面积预测方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21902637 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21902637 Country of ref document: EP Kind code of ref document: A1 |