An adaptive energy-saving diagnosis method based on coking big data and self-learning technology

Abstract

An adaptive energy-saving diagnosis method based on coking big data and self-learning technology solves the problem that the dynamic production process guided by static calculation and analysis results cannot adapt to the dynamic production process and continuously changing production rhythm of coking. This adaptive energy-saving diagnosis method can effectively monitor the complex production process of coking through the establishment of coking big data, realize online analysis and diagnosis of a series of substances such as dry coal, coke, coke oven gas, tar, ammonia, benzene and their correlation, and achieve the purpose of real-time dynamic, online processing and iterative upgrading of the production process and changing production rhythm of many equipment groups of coking by integrating machine self-learning technology. Thus, the data association of the whole coking plant is realized, the online collaborative monitoring and dynamic analysis of the production process of each equipment is realized, the energy-saving diagnosis results and dynamic feedback are given, and the adaptive collaborative energy-saving optimization of the coking plant based on the big data of the whole plant and the iterative upgrading of the system operation level are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial artificial intelligence for metallurgical automation, and specifically relates to an adaptive energy-saving diagnosis method based on coking big data and self-learning technology that can realize accurate measurement and online transmission of coking production data, effectively monitor the complex production process of coking, establish effective coking big data, realize dynamic energy-saving diagnosis of coking process, and utilize self-learning mechanism to continuously realize efficient iterative improvement of the operation process of coking system, thereby ensuring the adaptive optimization operation of coking system. Background Art

[0002] Coking plants take coke ovens as the core, and design a series of auxiliary production processes around the coke production process to form a complete coking production process. Coking plants use coal as raw material, and the main product is coke, and the by-products are coke oven gas and various purified additional products. Due to the complex coking production process, the dynamic calculation and online monitoring of the coking process are currently difficult. Coking plants are not only large energy consumers, but also difficult to monitor the energy consumption of complex smelting processes. The current solution is to conduct static design calculations and offline calculations around specific coking plant equipment or local process sections, and use static calculation analysis results to guide the dynamic production process. This method cannot adapt to the dynamic production process of coking and the continuously changing production rhythm. Therefore, on the one hand, it is known that there are many energy-saving and consumption-reducing nodes in the coking production process, and on the other hand, it is impossible to find a timely and comprehensive analysis and diagnosis method for coking plants, which makes the energy utilization rate of the coking production process low and the advancement of digital technology slow; therefore, it is necessary to improve the existing energy-saving diagnosis method for coking production. Summary of the invention

[0003] The present invention aims to solve the above problems and provides a method for realizing accurate measurement and online transmission of coking production data, effectively monitoring the complex production process of coking, establishing effective coking big data, realizing dynamic energy-saving diagnosis of coking process, and utilizing self-learning mechanism to continuously realize efficient iterative improvement of the operation process of coking system, thereby ensuring adaptive optimization operation of coking system.

[0004] The technical solution adopted by the present invention is: the adaptive energy-saving diagnosis method based on coking big data and self-learning technology includes the following steps:

[0005] Step 1: Establish a coking production big data generation model; combine the four characteristics of big data, namely, large volume, high speed, diversity and value, and through the effective application of online data transmission technology, data precision measurement technology, process principle technology and full coverage correlation technology, the coking production process will be transformed into effective industrial big data;

[0006] Step 2: Establish an iterative upgrade model for the coking machine self-learning system; the machine self-learning process based on the coking production big data resources includes a four-layer system that is interrelated; based on the coking plant's production equipment and the entire process system, the first layer is the big data layer, the second layer is the data processing layer, the third layer is the full process information association, training and equipment operation improvement layer, and the fourth layer is the judgment index and system model upgrade layer;

[0007] Step 3. Establish a coking plant adaptive collaborative optimization operation model; achieve green operation of the entire plant by setting reasonable ranges of key energy-saving parameters for the entire process; further form adaptive behavior guidelines for the entire coking plant by designing feedback, forecasting, and optimization algorithms combined with execution mechanisms to regulate energy-saving and efficient operation of the entire plant.

[0008] In the step 1, the supporting technologies used in the online data transmission technology include computer technology, information technology and real-time database system technology, thereby establishing a unified production information integration platform for the coking plant; the core part of the production information integration platform is its data center, which is established using network technology and IP real-time database system, and can realize the online dynamic intercommunication of various raw material data, process flow data, equipment operation process data, and PLC system and DCS system of the coking plant.

[0009] In the step 1, the main models corresponding to the data precision measurement technology are shown in formulas (1) to (5);

[0010] Continuity equation:

[0011]

[0012] Momentum equation:

[0013]

[0014] Turbulent momentum equation:

[0015]

[0016]

[0017]

[0018] Where: ρ is the density of gas working conditions;

[0019] u, v, w—velocity components of the gas in the pipe network along the three directions of the coordinate axis;

[0020] x, y and z—represent the three directions of the coordinate axis;

[0021] —Underlying variables in the control equation;

[0022] U—velocity vector of gas in the pipeline;

[0023] Γ—Variable The diffusion coefficient of

[0024] S—source term in the control equation;

[0025] μ—coal gas viscosity coefficient, dynamic viscosity;

[0026] μ t —turbulent viscosity coefficient;

[0027] p e —effective pressure in the turbulent momentum equation;

[0028] B x , B y , B z —Represents the components of body force in the x-, y-, and z-directions.

[0029] In the step 1, the expert system model in the process principle technology is established according to the equipment operation principle, and the model is shown in formula (6) to formula (8);

[0030] Product yield model:

[0031]

[0032] In the formula: i—represents category i products;

[0033] j—indicates equipment j that produces product category i;

[0034] P 产品,i — indicates the quantity of products of category i produced;

[0035] M 物料,i —Indicates the amount of materials used to produce category i products;

[0036] β i —Indicates the conversion coefficient of input materials corresponding to the production of category i products, statistical value;

[0037] The energy generated by coking is mainly secondary energy, including coke oven gas, steam, etc. Secondary energy is a valuable resource and should be recycled as efficiently as possible; the energy recovery model is:

[0038]

[0039] In the formula: k—represents type k of secondary energy;

[0040] j—indicates the equipment j that generates secondary energy;

[0041] E 能源产生,k —Indicates the amount of recovered secondary energy of type k;

[0042] M 物料,k —Indicates the amount of material consumed in producing type k of secondary energy;

[0043] α k —Indicates the conversion coefficient of input materials corresponding to type k of secondary energy, statistical value;

[0044] η j — represents the recovery efficiency of secondary energy k by device j;

[0045] Energy consumption model:

[0046]

[0047] In the formula: k—represents the k-type energy consumed;

[0048] j—indicates the equipment j that consumes energy;

[0049] E 能源消耗,k — represents the consumption of type k energy by type j equipment;

[0050] M 物料,k —Indicates the amount of material consumed by equipment of type j that consumes energy of type k;

[0051] α k —Indicates the conversion coefficient of the consumed materials corresponding to the k-type energy, statistical value;

[0052] η j —Represents the design margin of equipment j for energy consumption k, a statistical value.

[0053] The step 1, full coverage correlation technology, establishes a data correlation model for the whole process of the coking plant through the input-output model of materials and energy in each process area of ​​the coking plant combined with the concept of correlation, thereby forming the big data of coking production; in order to establish the correlation model of the whole coking plant and implement full coverage of the whole coking process, the coking production is divided into several areas;

[0054] The main material connection between the coking-four-car area and the dry coke quenching area, which are the core of coking, is reflected in the coke in the coke oven area being transported to the dry coke quenching area by the coke quenching car. In the process of establishing the material and energy input-output model of the coking area, considering that the flow rates of gas, air, electricity, water, flue gas and smoke can be directly measured, there is no need to establish a calculation model, while the remaining materials and energy need to establish models for calculation;

[0055] 1. The coking-four-car regional model is shown in equations (9) to (17);

[0056] Dry coal quality:

[0057]

[0058] Where: - the quality of dry coal in the coke oven;

[0059] ——The moisture content of the coal after drying during coal conditioning;

[0060] ——The amount of coal dried in the second generation of coal moisture conditioning;

[0061] Coke quality:

[0062]

[0063] Where: - the quality of coke produced in the coke oven;

[0064] ——coke yield of dry coal in the coke oven;

[0065] - the quality of dry coal in the coke oven;

[0066] Raw gas volume:

[0067]

[0068] Where: ——The volume of raw gas produced in the coke oven;

[0069] ——coke yield of dry coal in the coke oven;

[0070] ——density of raw gas in coke oven;

[0071] - the quality of dry coal in the coke oven;

[0072] Tar quality:

[0073]

[0074] Where: - the quality of tar produced in the coke oven;

[0075] ——The mass of tar produced in the coke oven as a percentage of the mass of dry coal;

[0076] - the quality of dry coal in the coke oven;

[0077] Tar residue quality:

[0078]

[0079] Where: - the quality of the fuel slag produced in the coke oven;

[0080] ——The percentage of tar residue mass produced in the coke oven to the mass of dry coal;

[0081] - the quality of tar produced in the coke oven;

[0082] Water Quality:

[0083]

[0084] Where: - the amount of water produced in the coke oven;

[0085] ——the amount of water produced in the coke oven as a percentage of the amount of coal;

[0086] ——Quality of coal blend entering the coke oven;

[0087]

[0088] Where: ——the amount of water produced in the coke oven as a percentage of the amount of coal;

[0089] ——moisture content of blended coal;

[0090] ——The mass of combined water in the coke oven as a percentage of the dry coal mass;

[0091] Ammonia Quality:

[0092]

[0093] Where: - the amount of ammonia produced in the coke oven;

[0094] - the amount of ammonia produced in the coke oven as a percentage of the dry coal;

[0095] ——Quality of coal blend entering the coke oven;

[0096] Benzene mass:

[0097]

[0098] Where: - the amount of benzene produced in the coke oven;

[0099] ——the amount of benzene produced in the coke oven as a percentage of the dry coal;

[0100] ——Quality of coal blend entering the coke oven;

[0101] 2. The dry quenching regional model is shown in equations (18) to (19);

[0102]

[0103] Where: ——the amount of coke powder in CDQ dedusting;

[0104] ——Coke powder rate in CDQ dust removal;

[0105] - the amount of coke produced in the coke oven;

[0106]

[0107] Where: ——the amount of coke in CDQ;

[0108] ——coke burnout rate in CDQ coke burnout;

[0109] - the amount of coke produced in the coke oven;

[0110] ——The amount of coke in CDQ dust removal.

[0111] Step 2: By inputting the historical big data collected by each device in the large system, the future operation status is predicted and the important status and attention indicators are predicted; in addition, according to the historical operation rules and current status of the specific equipment, the product output, quality, detection rate and equipment system status are predicted; the established machine self-learning model prediction and state prediction model are as follows:

[0112] Exponential smoothing forecasting algorithm:

[0113] S t =a×y t +(1-a)×S t-1 (20)

[0114] Where: S t represents the smoothed value at time t;

[0115] y t represents the actual value of time t;

[0116] S t-1 represents the smoothed value at time t-1;

[0117] a represents the smoothing coefficient, which ranges from [0,1].

[0118] The exponential smoothing prediction algorithm includes:

[0119] 1. Exponential smoothing algorithm: When there is no obvious trend change in the time series, exponential smoothing prediction can be used; its prediction formula is:

[0120] (y t+1 )′=a×y t +(1-a)×(y t )' (twenty one)

[0121] Where: (y t+1 )′ represents the predicted value of period t+1, that is, the smoothed value S of the current period (period t) t ;

[0122] y t represents the actual value of period t;

[0123] (y t )′ represents the forecast value of period t, that is, the smoothed value S of the previous period t-1 ;

[0124] 2. Secondary exponential smoothing algorithm: Based on the first exponential smoothing, the mathematical model of secondary exponential smoothing is:

[0125] Y t+T =a t +b t ×T (22)

[0126]

[0127]

[0128] Where: represents the quadratic exponential mean value of the tth period;

[0129] represents the exponential smoothing value of the tth period;

[0130] represents the quadratic exponential smoothing value of the t-1th period;

[0131] a represents the weighting coefficient (also called smoothing coefficient);

[0132] 3. Triple exponential smoothing algorithm: If the change of the time series shows a quadratic curve trend, the triple exponential smoothing method needs to be used for prediction; triple exponential smoothing is an additional smoothing based on quadratic exponential smoothing, and its mathematical model is:

[0133] Y t+T =a t +b t ×T+ct ×T 2 (25)

[0134]

[0135]

[0136]

[0137] Where: represents the triple exponential average of the tth period;

[0138] represents the quadratic exponential mean value of the tth period;

[0139] represents the exponential smoothing value of the tth period;

[0140] a represents the weighting coefficient.

[0141] In step 2, when there are many influencing parameters and links within the device or system, and the links are interrelated, a long short-term memory neural network algorithm is used; this method can help current decision-making by using historical information, and the long short-term memory network LSTM can well solve the problem of long-term dependence in the calculation process; the modeling process is as follows:

[0142] 1. Forget gate model: The first step of the long short-term memory network LSTM is to decide what information should be forgotten by the neuron. This is composed of a Sigmoid layer called the "forget gate layer". It inputs h t-1 and x t , then in C t-1 Each neuron state outputs a number between 0 and 1; "1" means "completely retained" and "0" means "completely forgotten". The mathematical model in the forget gate function is:

[0143] f t =σ(W f ×[h t-1 ,x t ]+b f ) (29)

[0144] Where: h t-1 —Indicates the state at the last moment;

[0145] x t —Indicates the value at the last moment;

[0146] f t — represents the forget gate function;

[0147] σ—represents the Sigmoid function;

[0148] f t —Indicates that the output state is 0 to 1;

[0149] a and b are coefficients;

[0150] 2. Input gate function: The input gate determines whether new information is placed in the cell state. It first passes through a tanh layer to obtain the representation of the current information. It also passes through a sigmoid layer to calculate which of the new information is important and which is not. It is then multiplied by the output of the tanh layer and added to the cell state. In short, this process weights the current new information into the cell state. After these operations, the cell state at the current moment can be obtained.

[0151] The mathematical model in the input gate function and activation function is:

[0152] i t =σ(W i ×[h t-1 ,x t ]+b i ) (30)

[0153]

[0154] Where: h t-1 —Indicates the state at the last moment;

[0155] x t —Indicates the value at the last moment;

[0156] σ—represents the forget gate function;

[0157] tanh — represents the activation function;

[0158] —Indicates the cell state at the current moment;

[0159] 3. Output gate function: The output gate determines the output state at this moment. It first converts the cell state at the current moment After a tanh layer excitation, and then by h t-1 and x t After a Sigmoid layer, the updated weights are obtained, and the two are multiplied; the result is the current cell state weighted by tanh excitation to obtain the current state;

[0160]

[0161] Where: C t —Indicates that the current cell state is weighted after tanh excitation to obtain the current state;

[0162] f t — indicates the result of the forget gate function;

[0163] C t-1 — represents the state of the cell at the previous time after the weighted state after tanh excitation;

[0164] — indicates the cell state at the current moment.

[0165] In the third step, all parameters corresponding to the industrial big data should conform to the professionalism of the equipment group operation and operation, so the big data of the coking process should have certain attributes and rules for easy classification; the corresponding parameters cannot have infinite value changes, and there should be reasonable upper and lower limits of parameters that conform to the operation principle of the equipment group. Based on this rule, a coking adaptive energy-saving diagnosis and optimization mechanism is established;

[0166] 1. Judgment of upper and lower limits of all data in the big data center:

[0167] Corresponding to all parameters in each area of ​​the entire coking process, the upper and lower limits of parameters are designed according to the equipment process principles, and relevant models are established to guide operations;

[0168] Parameter range model:

[0169]

[0170] In the formula: i—represents the i-th production equipment in the coking plant;

[0171] j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27;

[0172] X ij — indicates the data corresponding to device i in production area j transmitted by the data center;

[0173] —Indicates the minimum value allowed for this parameter based on the equipment process principle;

[0174] —Indicates the maximum value of the parameter according to the equipment process principle;

[0175] Equipment operation rule model corresponding to the parameters:

[0176]

[0177] 2. Use parameter data to calculate indicators, determine upper and lower limits of indicators, and guide operations:

[0178] The index η is determined by the index X ij , Y ij , ij ,ij Calculated, through the relationship between the index and the parameter variable, find out the weight influence size and sort it, and improve the corresponding operations in turn;

[0179]

[0180] In the formula: i—represents the i-th production equipment in the coking plant;

[0181] j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27;

[0182] η k —Indicates the kth indicator of concern in the coking production process;

[0183] X ij , Y ij , ij , ij —Indicates the production parameters obtained by the coking plant data center;

[0184] α, β, γ, — represents parameter variable X ij , Y ij , ij , ij The impact weight on the indicator η;

[0185] 3. Calculate the whole plant indicators, seek the best feedback, and guide execution:

[0186]

[0187] Where: η q,n —Indicates the nth indicator of concern at the coking plant level;

[0188] By establishing effective coking big data and utilizing machine self-learning methods combined with rule setting and feedback mechanisms, the adaptive collaborative energy-saving optimization operation and iterative upgrade of the coking system can be achieved.

[0189] Beneficial effects of the present invention: The adaptive energy-saving diagnosis method based on coking big data and self-learning technology establishes data association of the whole process of the coking plant through the process principle model of each equipment node, the input-output model of materials and energy in each process area, and the concept of correlation, thereby forming coking production big data, realizing accurate measurement and online transmission of coking production data; the establishment of coking big data can effectively monitor the complex production process of coking, and then conduct online analysis and diagnosis of a series of substances such as dry coal, coke, coke oven gas, tar, ammonia, benzene and their correlation. In addition, by integrating machine self-learning technology, the real-time dynamic, online processing and iterative upgrading of the production process and production rhythm of many equipment groups of coking can be realized; the internal benchmarking and industry benchmarking of the production process of each equipment of coking can be realized, and the self-learning iterative upgrading of equipment operation optimization of coking machines can be formed. The self-diagnosis model and the automatic feedback model are used to establish the automatic association of each area of ​​coking production, and combined with the online feedback diagnosis of the system, the purpose of establishing a coking adaptive collaborative optimization operation mechanism is achieved. The application of this adaptive energy-saving diagnosis method can achieve data association for the entire coking plant, online collaborative monitoring and dynamic analysis of the production processes of each equipment, provide dynamic feedback on the energy-saving diagnosis results, and realize adaptive collaborative energy-saving optimization of the coking plant based on the big data of the entire plant and iterative upgrade of the system operation level; improve the level of intelligence in this field, and provide reference and reference for industrial enterprises, which can play a huge role in promoting the rapid development of China's industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0190] Figure 1 It is the production process flow chart of the coking plant.

[0191] Figure 2 It is a schematic diagram of the existing coking energy-saving diagnosis working model and existing problems.

[0192] Figure 3 It is the big data characteristics and the corresponding supporting technology of the present invention.

[0193] Figure 4 It is a simplified diagram of the principle and function of online data transmission of the present invention.

[0194] Figure 5 It is a functional diagram of the data precision measurement system function and process principle system of the present invention.

[0195] Figure 6 It is a coking plant-wide association coverage correlation model and a functional schematic diagram of the present invention.

[0196] Figure 7 It is a schematic diagram of the input-output situation and correlation of the coking and coking area and the coke quenching area of ​​the present invention.

[0197] Figure 8 It is a machine self-learning flow chart of the entire coking process based on big data resources.

[0198] Fig. 9 It is a coking adaptive energy-saving diagnosis optimization mechanism method and a flow chart of the present invention. DETAILED DESCRIPTION

[0199] Industrial artificial intelligence has been widely mentioned. Industrial big data is not only the supporting technology of industrial artificial intelligence, but also one of the most effective ways for industrial enterprises to achieve intelligent manufacturing. However, there is no clear understanding of how to form effective industrial big data and how to make industrial artificial intelligence play a specific role in industrial enterprises.

[0200] Typical production process of coking plant (such as Figure 1 As shown in the figure, there are a series of main production processes including thawing, coal storage, coal extraction, coal blending, coal tower, coke oven, dry quenching, coke screening, gas-liquid separation, condensation, desulfurization, deamination, debenzene, hydrogen production, etc. The current main modes of energy conservation and consumption reduction in coking production are as follows: Figure 2 As shown in the figure, it can be seen that the lack of accurate process data, the lack of effective accounting models for the main coking equipment and purification processes, and the lack of effective connections and feedback functions among the many equipment in the coking plant have led to the inability to effectively carry out coking energy-saving diagnosis work.

[0201] The method of the present invention establishes data association of the whole process of the coking plant by combining the process principle model of each equipment node, the input-output model of materials and energy in each process area with the concept of correlation, thereby forming big data of coking production. By establishing a self-learning mechanism, the operation process of the coking system can be continuously improved by efficient iteration. The self-adaptive optimization operation mechanism of the coking plant is established through the self-diagnosis model and the automatic feedback model. On this basis, the efficient and energy-saving diagnosis of the coking production process based on big data, machine self-learning and adaptive collaboration is realized, and the intelligence level in this field is improved.

[0202] The specific steps of the present invention are described in detail. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology includes:

[0203] Step 1: Establish a coking production big data generation model.

[0204] At present, there are many ideas for the generation and use of big data in industrial enterprises, but they are not applicable in the coking site. The actual reasons are: first, the data is inaccurate, and the enterprise measurement data cannot accurately reflect the real production process; second, the big data formed by the accumulation of inaccurate data cannot guide the production of enterprises, although the workload is large, it is meaningless; third, each type of industrial enterprise has its own unique production process characteristics, and the general concept of big data establishment is not applicable to the application of industrial enterprises. Therefore, the effective establishment of coking production big data is both a difficult point in the industry and has great practical significance.

[0205] The present invention directly solves the above difficulties and is designed in combination with the "4V" characteristics of big data. Figure 3 It can be seen that, measured by the four characteristics of big data, the practicality of current industry big data projects is very limited. To this end, the present invention proposes four technologies as basic support to ensure the effective formation of industrial big data. The corresponding relationship is: first, to ensure the data online transmission technology and full coverage association technology of the Volume attribute of big data; second, to ensure the data online transmission technology and precise measurement technology of the Velocity attribute of big data; third, to ensure the process principle technology and full coverage association technology of the Variety attribute of big data; fourth, to ensure the precise measurement technology and process principle technology of the Value attribute of big data. Through the effective application of these supporting technologies, the coking production process will truly form effective industrial big data.

[0206] 1. Online data transmission technology

[0207] The implementation process of the data online transmission function of the present invention is as follows Figure 4 As shown, the supporting technologies used include advanced computer technology, information technology and real-time database system technology to establish a unified production information integration platform for coking plants; the core part of the platform is its data center, which is established using advanced network technology and IP real-time database system, and can realize online dynamic intercommunication of various raw material data, process flow data, equipment operation process data, PLC system, DCS system, etc. in the coking plant. Through the data online transmission system of the patent of this invention, various data of the coking plant can be transmitted in real time, various parts of the production process can be interconnected, information can be shared and highly integrated, and the integration of the management layer and the control layer can be realized. This part is the basis for ensuring the smooth transmission of coking big data.

[0208] 2. Data precision measurement technology and process principle technology

[0209] The schematic diagram of the data precision measurement system function and process principle system function of the present invention is as follows: Figure 5 The actual working conditions at the enterprise site are the biggest problem affecting the accuracy of the metering system data. The present invention solves this problem by establishing a corresponding unified model for the working conditions of insufficient straight pipe sections, small flow conditions, cold and wet medium conditions, and wide range conditions.

[0210] ① The main models corresponding to data precision measurement technology are shown in equations (1) to (5);

[0211] Continuity equation:

[0212]

[0213] Momentum equation:

[0214]

[0215] Turbulent momentum equation:

[0216]

[0217]

[0218]

[0219] Where: ρ is the density of gas working conditions;

[0220] u, v, w—velocity components of the gas in the pipe network along the three directions of the coordinate axis;

[0221] x, y and z—represent the three directions of the coordinate axis;

[0222] —Underlying variables in the control equation;

[0223] U—velocity vector of gas in the pipeline;

[0224] Γ—Variable The diffusion coefficient of

[0225] S—source term in the control equation;

[0226] μ—coal gas viscosity coefficient, dynamic viscosity;

[0227] μ t —turbulent viscosity coefficient;

[0228] p e —effective pressure in the turbulent momentum equation;

[0229] B x , B y , B z —Represents the components of body force in the x-, y-, and z-directions.

[0230] ② The expert system model in the process principle technology is established based on the equipment operation principle, and the model is shown in equations (6) to (8). Due to the large number of coking plant equipment groups and their various types, a highly condensed model form will be shown here.

[0231] Product yield model:

[0232]

[0233] In the formula: i—represents category i products;

[0234] j—indicates equipment j that produces product category i;

[0235] P 产品,i — indicates the quantity of products of category i produced;

[0236] M 物料,i —Indicates the amount of materials used to produce category i products;

[0237] β i —Indicates the conversion coefficient of input materials corresponding to the production of type i products, statistical value.

[0238] The energy generated by coking is mainly secondary energy, including coke oven gas, steam, etc. Secondary energy is a valuable resource and should be recycled as efficiently as possible; the energy recovery model is:

[0239]

[0240] In the formula: k—represents type k of secondary energy;

[0241] j—indicates the equipment j that generates secondary energy;

[0242] E 能源产生,k —Indicates the amount of recovered secondary energy of type k;

[0243] M 物料,k —Indicates the amount of material consumed in producing type k of secondary energy;

[0244] α k —Indicates the conversion coefficient of input materials corresponding to type k of secondary energy, statistical value;

[0245] η j — represents the recovery efficiency of device j for secondary energy k.

[0246] Energy consumption model:

[0247]

[0248] In the formula: k—represents the k-type energy consumed;

[0249] j—indicates the equipment j that consumes energy;

[0250] E 能源消耗,k — represents the consumption of type k energy by type j equipment;

[0251] M 物料,k —Indicates the amount of material consumed by equipment of type j that consumes energy of type k;

[0252] α k —Indicates the conversion coefficient of the consumed materials corresponding to the k-type energy, statistical value;

[0253] η j —Represents the design margin of equipment j for energy consumption k, a statistical value.

[0254] 3. Full coverage of related technologies

[0255] The schematic diagram of the plant-wide coverage association function of the present invention is as follows: Figure 6 As shown in the figure. By combining the input-output model of materials and energy in each process area of ​​coking with the concept of correlation, a data association model for the entire process of the coking plant is established to form the big data of coking production.

[0256] In order to establish a correlation model for the entire coking plant and implement full coverage of the entire coking process, the coking production is divided into 27 areas after analysis, as shown in the following table.

[0257]

[0258]

[0259] The data association model of the whole coking plant includes two main parts: input-output model and correlation model. The above 27 areas are processed by combining these two models, and the patterns are basically the same. For the sake of convenience, the core coking-four-car area (the 8th area) and dry quenching area (the 9th area) are taken as examples. The input-output and correlation of materials and energy in these two areas are as follows: Figure 7 As shown, the main material connection between the two areas is that the coke from the coke oven area is transported to the dry quenching area by quenching cars.

[0260] In the process of establishing the material and energy input-output model of the coking area, it is considered that the flow rates of coal gas (coke oven gas COG, blast furnace gas BFG), air, electricity, water, flue gas and smoke dust can be directly measured and there is no need to establish a calculation model. However, models need to be established for calculation of the remaining materials and energy.

[0261] ① The coking-four-car regional model is shown in equations (9) to (17);

[0262] Dry coal quality:

[0263]

[0264] Where: - the quality of dry coal in the coke oven;

[0265] ——The moisture content of the coal after drying during coal conditioning;

[0266] ——The amount of coal dried during the second generation of coal moisture conditioning.

[0267] Coke quality:

[0268]

[0269] Where: - the quality of coke produced in the coke oven;

[0270] ——coke yield of dry coal in the coke oven;

[0271] ——The quality of dry coal in the coke oven.

[0272] Raw gas volume:

[0273]

[0274] Where: ——The volume of raw gas produced in the coke oven;

[0275] ——coke yield of dry coal in the coke oven;

[0276] ——density of raw gas in coke oven;

[0277] ——The quality of dry coal in the coke oven.

[0278] Tar quality:

[0279]

[0280] Where: - the quality of tar produced in the coke oven;

[0281] ——The mass of tar produced in the coke oven as a percentage of the mass of dry coal;

[0282] ——The quality of dry coal in the coke oven.

[0283] Tar residue quality:

[0284]

[0285] Where: - the quality of the fuel slag produced in the coke oven;

[0286] ——The percentage of tar residue mass produced in the coke oven to the mass of dry coal;

[0287] ——The quality of tar produced in the coke oven.

[0288] Water Quality:

[0289]

[0290] Where: - the amount of water produced in the coke oven;

[0291] ——the amount of water produced in the coke oven as a percentage of the amount of coal;

[0292] ——The quality of the coal entering the coke oven.

[0293]

[0294] Where: ——the amount of water produced in the coke oven as a percentage of the amount of coal;

[0295] ——moisture content of blended coal;

[0296] ——The percentage of combined water mass in the coke oven to the dry coal mass.

[0297] Ammonia Quality:

[0298]

[0299] Where: - the amount of ammonia produced in the coke oven;

[0300] - the amount of ammonia produced in the coke oven as a percentage of the dry coal;

[0301] ——The quality of the coal entering the coke oven.

[0302] Benzene mass:

[0303]

[0304] Where: - the amount of benzene produced in the coke oven;

[0305] ——the amount of benzene produced in the coke oven as a percentage of the dry coal;

[0306] ——The quality of the coal entering the coke oven.

[0307] ② The dry quenching area model is shown in equations (18) to (19);

[0308]

[0309] Where: ——the amount of coke powder in CDQ dedusting;

[0310] ——Coke powder rate in CDQ dust removal;

[0311] ——The amount of coke produced in the coke oven.

[0312]

[0313] Where: ——the amount of coke in CDQ;

[0314] ——coke burnout rate in CDQ coke burnout;

[0315] - the amount of coke produced in the coke oven;

[0316] ——The amount of coke in CDQ dust removal.

[0317] Step 2: Establish an iterative upgrade model for the coking machine self-learning system.

[0318] Through the above-mentioned online, dynamic, accurate data, process accounting, full-process data association and other functions, the basic work for the generation of big data in the coking industry has been done. Next, effective industrial big data is generated for system use, which also greatly promotes the dynamic real-time calculation and accuracy of the current industrial principle model. The present invention has designed a machine self-learning function based on big data to realize automatic iterative upgrades of the system and form a new correlation between the entire coking process. Machine self-learning based on big data will have a very wide range of applications, but the purpose of the present invention is to solve the problem of energy-saving diagnosis. The process of the machine self-learning mode designed for this purpose is as follows: Figure 8 shown.

[0319] The present invention is based on the machine self-learning process of big data resources and consists of a four-layer system that is interrelated. Based on the production equipment and the whole process system of the coking plant, Figure 8 From bottom to top, the first layer is the big data layer; the second layer is the data processing layer; the third layer is the full-process information association, training and equipment operation layer; the fourth layer is the judgment index and system model upgrade layer. Because industrial big data has its inherent process attributes, it is different from the general big data in other fields of society. The data generated by the industry has actual physical meaning and represents certain characteristics. For this reason, the second layer is used to process data and analyze and process a large amount of data through the material and energy balance laws of the industrial production system. The third layer is used to train and guide the optimization of industrial systems. Through the human-machine perspective function, each enterprise can formulate various production indicators of concern to evaluate the effect of optimized operation improvement. If the improvement effect is not ideal, it will be fed back to the second layer for adjustment; if the enterprise production process evaluation indicators are actually improved, the established system model group can be further improved, and the system can be constrained at a higher level and more strictly to improve the effectiveness of the system.

[0320] By inputting the historical big data collected by each device in the large system, the future operation status is predicted and important situations and indicators of concern are predicted; in addition, according to the historical operation rules of specific equipment and the current situation, the product output, quality, detection rate, and equipment system status are predicted. The machine self-learning model prediction and state prediction model established are as follows:

[0321] 1. Exponential smoothing forecasting algorithm

[0322] S t =a×y t +(1-a)×S t-1 (20)

[0323] Where: S t represents the smoothed value at time t;

[0324] y t represents the actual value of time t;

[0325] S t-1 represents the smoothed value at time t-1;

[0326] a represents the smoothing coefficient, which ranges from [0,1].

[0327] ①One-step exponential smoothing algorithm

[0328] When there is no obvious trend change in the time series, an exponential smoothing forecast can be used. The forecast formula is:

[0329] (y t+1 )′=a×y t +(1-a)×(y t )' (twenty one)

[0330] Where: (y t+1 )′ represents the predicted value of period t+1, that is, the smoothed value S of the current period (period t) t ;

[0331] y t represents the actual value of period t;

[0332] (y t )′ represents the forecast value of period t, that is, the smoothed value S of the previous period t-1 .

[0333] ② Quadratic exponential smoothing algorithm

[0334] On the basis of the first exponential smoothing, the mathematical model of the second exponential smoothing is obtained as follows:

[0335] Y t+T =a t +b t ×T (22)

[0336]

[0337]

[0338] Where: represents the quadratic exponential mean value of the tth period;

[0339] represents the exponential smoothing value of the tth period;

[0340] represents the quadratic exponential smoothing value of the t-1th period;

[0341] a represents the weighting coefficient (also called the smoothing coefficient).

[0342] ③Triple exponential smoothing algorithm

[0343] If the change of the time series shows a quadratic curve trend, it is necessary to use the triple exponential smoothing method for prediction; triple exponential smoothing is an additional smoothing based on quadratic exponential smoothing, and its mathematical model is:

[0344] Y t+T =a t +b t ×T+c t ×T 2 (25)

[0345]

[0346]

[0347]

[0348] Where: represents the triple exponential average of the tth period;

[0349] represents the quadratic exponential mean value of the tth period;

[0350] represents the exponential smoothing value of the tth period;

[0351] a represents the weighting coefficient (also called the smoothing coefficient).

[0352] 2. Long Short-Term Memory Neural Network Algorithm

[0353] When there are many influencing parameters and links within the equipment or system, and the links are interrelated, the long short-term memory neural network algorithm is used; this method can help current decision-making by using historical information, and the long short-term memory network LSTM can well solve the problem of long-term dependence in the calculation process; the model construction process is as follows:

[0354] ① Forget gate model: The first step of the long short-term memory network LSTM is to decide what information should be forgotten by the neuron. This is composed of a Sigmoid layer called the "forget gate layer"; it inputs h t-1 and x t , then in C t-1 Each neuron state outputs a number between 0 and 1; "1" means "completely retained" and "0" means "completely forgotten". The mathematical model in the forget gate function is:

[0355] f t =σ(W f ×[h t-1 ,x t ]+b f ) (29)

[0356] Where: h t-1 —Indicates the state at the last moment;

[0357] x t —Indicates the value at the last moment;

[0358] f t — represents the forget gate function;

[0359] σ—represents the Sigmoid function;

[0360] f t —Indicates that the output state is 0 to 1;

[0361] a and b are coefficients.

[0362] ② Input gate function: The input gate determines whether new information is placed in the cell state. It first passes through a tanh layer to obtain the representation of the current information. It also passes through a sigmoid layer to calculate which of the new information is important and which is not. It is then multiplied by the output of the tanh layer and added to the cell state. In short, this process is to weight the current new information into the cell state. After these operations, the cell state at the current moment can be obtained.

[0363] The mathematical model in the input gate function and activation function is:

[0364] i t =σ(W i ×[ht-1 ,x t ]+b i ) (30)

[0365]

[0366] Where: h t-1 —Indicates the state at the last moment;

[0367] x t —Indicates the value at the last moment;

[0368] σ—represents the forget gate function;

[0369] tanh — represents the activation function;

[0370] —Indicates the cell state at the current moment.

[0371] ③ Output gate function: The output gate determines the output state at this moment. It first converts the cell state at the current moment After a tanh layer excitation, and then by h t-1 and x t After a Sigmoid layer, the updated weights are obtained, and the two are multiplied; the result is the current cell state weighted by tanh excitation to obtain the current state;

[0372]

[0373] Where: C t —Indicates that the current cell state is weighted after tanh excitation to obtain the current state;

[0374] f t — indicates the result of the forget gate function;

[0375] C t-1 — represents the state of the cell at the previous time after the weighted state after tanh excitation;

[0376] — indicates the cell state at the current moment.

[0377] When industrial big data develops to a certain stage, it can further promote the process of industrial artificial intelligence and make intelligent manufacturing a truly achievable goal in the industrial field. The integration of machine self-learning technology can realize the real-time dynamic, online processing and iterative upgrade of the production process and production rhythm of many equipment groups in the coking plant.

[0378] Step 3: Establish a coking adaptive collaborative optimization operation model.

[0379] The adaptive energy-saving diagnosis and optimization mechanism of the coking system needs to be based on the above work, and realize the green operation of the whole plant by setting reasonable ranges of key energy-saving parameters in the whole process; further, by designing feedback, forecasting, and optimization algorithms and combining them with execution mechanisms, an adaptive behavior code for the whole coking plant is formed to regulate the energy-saving and efficient operation of the whole plant.

[0380] Big data in the industrial field, especially the process principles and internal laws, are different from the general big data in other fields (such as the consumer field). All parameters corresponding to the big data in the industrial field should conform to the professionalism of the operation and operation of the equipment group. Therefore, the big data of the coking process in the present invention should be classifiable, have certain properties and laws, and the corresponding various parameters cannot have infinite value changes. There should be reasonable upper and lower limits of parameters that conform to the operating principles of the equipment group. Based on this law, a coking adaptive energy-saving diagnosis and optimization mechanism is established (the schematic diagram of the method and process, such as Fig. 9 shown):

[0381] 1. Judgment of upper and lower limits of all data in the big data center:

[0382] Corresponding to all parameters in 27 areas of the whole coking process, the upper and lower limits of parameters are designed according to the equipment process principle and the relevant models are established to guide the operation. Due to the huge number of parameters in the whole coking process, for the convenience of expression, the present invention introduces a highly condensed general model form, and the subsequent indicator series also adopts the same processing method.

[0383] Parameter range model:

[0384]

[0385] In the formula: i—represents the i-th production equipment in the coking plant;

[0386] j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27;

[0387] X ij — indicates the data corresponding to device i in production area j transmitted by the data center;

[0388] —Indicates the minimum value allowed for this parameter based on the equipment process principle;

[0389] —Indicates the maximum value of the parameter based on the equipment process principle.

[0390] Equipment operation rule model corresponding to the parameters:

[0391]

[0392] 2. Use parameter data to calculate indicators, determine upper and lower limits of indicators, and guide operations:

[0393] The index η is determined by the index X ij , Y ij , ij , ij Calculated, through the relationship between the index and the parameter variable, find out the weight influence size and sort it, and improve the corresponding operations in turn;

[0394]

[0395] In the formula: i—represents the i-th production equipment in the coking plant;

[0396] j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27;

[0397] η k —Indicates the kth indicator of concern in the coking production process;

[0398] X ij , Y ij , ij , ij —Indicates the production parameters obtained by the coking plant data center;

[0399] α, β, γ, — represents parameter variable X ij , Y ij , ij , ij The influence weight on the indicator η.

[0400] 3. Calculate the whole plant indicators, seek the best feedback, and guide execution:

[0401]

[0402] Where: η q,n — represents the nth index of concern at the coking plant level. The physical meanings of other variables and superscripts are the same as those in equations (33) and (34), and will not be repeated here.

[0403] By establishing effective coking big data, using machine self-learning methods combined with rule setting and feedback mechanisms, the coking system can achieve adaptive collaborative energy-saving optimization operation and iterative upgrades. This can provide reference and reference for industrial enterprises and play a huge role in promoting the rapid development and intelligence of China's industry.

Claims

1. An adaptive energy-saving diagnosis method based on coking big data and self-learning technology, characterized in that: The steps include: Step 1: Establish a coking production big data generation model; combine the four characteristics of big data, namely, large volume, high speed, diversity and value, and through the effective application of online data transmission technology, data precision measurement technology, process principle technology and full coverage correlation technology, the coking production process will be transformed into effective industrial big data; Step 2: Establish an iterative upgrade model for the coking machine self-learning system; the machine self-learning process based on the coking production big data resources includes a four-layer system that is interrelated; based on the coking plant's production equipment and the entire process system, the first layer is the big data layer, the second layer is the data processing layer, the third layer is the full process information association, training and equipment operation improvement layer, and the fourth layer is the judgment index and system model upgrade layer; Step 3: Establish a coking adaptive collaborative optimization operation model; achieve green operation of the whole plant by setting a reasonable range of key energy-saving parameters for the whole process; further form an adaptive behavior code for the whole coking plant by designing feedback, forecasting, and optimization algorithms combined with execution mechanisms to regulate the energy-saving and efficient operation of the whole plant; In the third step, all parameters corresponding to the industrial big data should conform to the professionalism of the equipment group operation and operation, so the big data of the coking process should have certain attributes and rules for easy classification; the corresponding parameters cannot have infinite value changes, and there should be reasonable upper and lower limits of parameters that conform to the operation principle of the equipment group. Based on this rule, a coking adaptive energy-saving diagnosis and optimization mechanism is established; ① Judgment of upper and lower limits of all data in the big data center: Corresponding to all parameters in each area of ​​the entire coking process, the upper and lower limits of parameters are designed according to the equipment process principles, and relevant models are established to guide operations; Parameter range model: In the formula: i—represents the i-th production equipment in the coking plant; j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27; X ij — indicates the data corresponding to device i in production area j transmitted by the data center; —Indicates the minimum value allowed for this parameter based on the equipment process principle; —Indicates the maximum value of the parameter according to the equipment process principle; Equipment operation rule model corresponding to the parameters: ② Use parameter data to calculate indicators, determine upper and lower limits of indicators, and guide operations: The index η is determined by the index X ij , Y ij , ij , ij Calculated, through the relationship between the index and the parameter variable, find out the weight influence size and sort it, and improve the corresponding operations in turn; In the formula: i—represents the i-th production equipment in the coking plant; j—represents the jth production area of ​​the coking plant, j=1,2,3,…,27; η k —Indicates the kth indicator of concern in the coking production process; X ij , Y ij , ij , ij —Indicates the production parameters obtained by the coking plant data center; α, β, γ, — represents parameter variable X ij , Y ij , ij , ij The impact weight on the indicator η; ③ Calculate the whole factory index, seek the best feedback and guide the execution: Where: η q,n —Indicates the nth indicator of concern at the coking plant level; By establishing effective coking big data and utilizing machine self-learning methods combined with rule setting and feedback mechanisms, the adaptive collaborative energy-saving optimization operation and iterative upgrade of the coking system can be achieved.

2. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: In the step 1, the supporting technologies used in the online data transmission technology include computer technology, information technology and real-time database system technology, thereby establishing a unified production information integration platform for the coking plant; the core part of the production information integration platform is its data center, which is established using network technology and IP real-time database system, and can realize the online dynamic intercommunication of various raw material data, process flow data, equipment operation process data, and PLC system and DCS system of the coking plant.

3. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: In the step 1, the main models corresponding to the data precision measurement technology are shown in formulas (1) to (5); Continuity equation: Momentum equation: Turbulent momentum equation: Where: ρ is the density of gas working conditions; u, v, w—velocity components of the gas in the pipe network along the three directions of the coordinate axis; x, y and z—represent the three directions of the coordinate axis; —Underlying variables in the control equation; U—velocity vector of gas in the pipeline; Γ—Variable The diffusion coefficient of S—source term in the control equation; μ—coal gas viscosity coefficient, dynamic viscosity; μ t —turbulent viscosity coefficient; p e —effective pressure in the turbulent momentum equation; B x , B y , B z —Represents the components of body force in the x-, y-, and z-directions.

4. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: In the step 1, the expert system model in the process principle technology is established according to the equipment operation principle, and the model is shown in formula (6) to formula (8); Product yield model: In the formula: i—represents category i products; j—indicates equipment j that produces product category i; P 产品,i — represents the quantity of products of category i produced; M 物料,i —Indicates the amount of materials used to produce category i products; β i —Indicates the conversion coefficient of input materials corresponding to the production of category i products, statistical value; The energy generated by coking is mainly secondary energy, including coke oven gas and steam. Secondary energy is a valuable resource and should be recycled as efficiently as possible. The energy recovery model is: In the formula: k—represents type k of secondary energy; j—indicates the equipment j that generates secondary energy; E 能源产生,k —Indicates the amount of recovered secondary energy of type k; M 物料,k —Indicates the amount of material consumed in producing type k of secondary energy; α k —Indicates the conversion coefficient of input materials corresponding to type k of secondary energy, statistical value; η j — represents the recovery efficiency of secondary energy k by device j; Energy consumption model: In the formula: k—represents the k-type energy consumed; j—indicates the equipment j that consumes energy; E 能源消耗,k — represents the consumption of type k energy by type j equipment; M 物料,k —Indicates the amount of material consumed by equipment of type j that consumes energy of type k; α k —Indicates the conversion coefficient of the consumed materials corresponding to the k-type energy, statistical value; η j —Represents the design margin of equipment j for energy consumption k, a statistical value.

5. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: The step 1, full coverage correlation technology, establishes a data correlation model for the whole process of the coking plant through the input-output model of materials and energy in each process area of ​​the coking plant combined with the concept of correlation, thereby forming the big data of coking production; in order to establish the correlation model of the whole coking plant and implement full coverage of the whole coking process, the coking production is divided into several areas; The main material connection between the coking-four-car area and the dry coke quenching area, which are the core of coking, is reflected in the coke in the coke oven area being transported to the dry coke quenching area by the coke quenching car. In the process of establishing the material and energy input-output model of the coking area, considering that the flow rates of gas, air, electricity, water, flue gas and smoke can be directly measured, there is no need to establish a calculation model, while the remaining materials and energy need to establish models for calculation; ① The coking-four-car regional model is shown in equations (9) to (17); Dry coal quality: Where: - the quality of dry coal in the coke oven; ——The moisture content of the coal after drying during coal conditioning; ——The amount of coal dried in the second generation of coal moisture conditioning; Coke quality: Where: - the quality of coke produced in the coke oven; ——coke yield of dry coal in the coke oven; - the quality of dry coal in the coke oven; Raw gas volume: Where: ——The volume of raw gas produced in the coke oven; ——coke yield of dry coal in the coke oven; ——density of raw gas in coke oven; - the quality of dry coal in the coke oven; Tar quality: Where: - the quality of tar produced in the coke oven; ——The mass of tar produced in the coke oven as a percentage of the mass of dry coal; - the quality of dry coal in the coke oven; Tar residue quality: Where: - the quality of the fuel slag produced in the coke oven; ——The percentage of tar residue mass produced in the coke oven to the mass of dry coal; - the quality of tar produced in the coke oven; Water Quality: Where: - the amount of water produced in the coke oven; ——the amount of water produced in the coke oven as a percentage of the amount of coal; ——Quality of coal blend entering the coke oven; Where: ——the amount of water produced in the coke oven as a percentage of the amount of coal; ——moisture content of blended coal; ——The mass of combined water in the coke oven as a percentage of the dry coal mass; Ammonia Quality: Where: - the amount of ammonia produced in the coke oven; - the amount of ammonia produced in the coke oven as a percentage of the dry coal; ——Quality of coal blend entering the coke oven; Benzene mass: Where: - the amount of benzene produced in the coke oven; ——the amount of benzene produced in the coke oven as a percentage of the dry coal; ——Quality of coal blend entering the coke oven; ② The dry quenching area model is shown in equations (18) to (19); Where: ——the amount of coke powder in CDQ dedusting; ——Coke powder rate in CDQ dust removal; - the amount of coke produced in the coke oven; Where: ——the amount of coke in CDQ; ——coke burnout rate in CDQ coke burnout; - the amount of coke produced in the coke oven; ——The amount of coke in CDQ dust removal.

6. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: Step 2: By inputting the historical big data collected by each device in the large system, the future operation status is predicted and the important status and attention indicators are predicted; in addition, according to the historical operation rules and current status of the specific equipment, the product output, quality, detection rate and equipment system status are predicted; the established machine self-learning model prediction and state prediction model are as follows: Exponential smoothing forecasting algorithm: S t =a×y t +(1-a)×S t-1 (20) Where: S t represents the smoothed value at time t; y t represents the actual value of time t; S t-1 represents the smoothed value at time t-1; a represents the smoothing coefficient, which ranges from [0,1].

7. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 6 is characterized by: The exponential smoothing prediction algorithm includes: ① Single exponential smoothing algorithm: When there is no obvious trend change in the time series, single exponential smoothing prediction can be used; its prediction formula is: (and t+1 )′=a×y t +(1-a)×(y t )′ (21) Where: (y t+1 )′ represents the predicted value of period t+1, that is, the smoothed value S of the current period, that is, period t t ; y t represents the actual value of period t; (y t )′ represents the forecast value of period t, that is, the smoothed value S of the previous period t-1 ; ② Secondary exponential smoothing algorithm: Based on the first exponential smoothing, the mathematical model of secondary exponential smoothing is: AND t+T =a t +b t ×T (22) Where: represents the quadratic exponential mean value of the tth period; represents the exponential smoothing value of the tth period; represents the quadratic exponential smoothing value of the t-1th period; a represents the weighting coefficient, also called the smoothing coefficient; ③Triple exponential smoothing algorithm; if the change of the time series shows a quadratic curve trend, the triple exponential smoothing method needs to be used for prediction; triple exponential smoothing is an additional smoothing based on quadratic exponential smoothing, and its mathematical model is: Y t+T =a t +b t ×T+c t ×T 2 (25) Where: represents the triple exponential average of the tth period; represents the quadratic exponential mean value of the tth period; represents the exponential smoothing value of the tth period; a represents the weighting coefficient.

8. The adaptive energy-saving diagnosis method based on coking big data and self-learning technology according to claim 1 is characterized in that: In the second step, when there are many influencing parameters and links in the equipment or system, and the links are interrelated, the long short-term memory neural network algorithm is used; this method can help current decision-making by using historical information, and the long short-term memory network LSTM can well solve the problem of long-term dependence in the calculation process; the modeling process is as follows: ① Forget gate model: The first step of the long short-term memory network LSTM is to decide what information should be forgotten by the neuron. This is composed of a Sigmoid layer called the "forget gate layer". It inputs h t-1 and x t , then in C t-1 Each neuron state outputs a number between 0 and 1; "1" means "completely retained" and "0" means "completely forgotten". The mathematical model in the forget gate function is: f t =σ(W f ×[h t-1 ,x t ]+b f ) (29) Where: h t-1 —Indicates the state at the last moment; x t —Indicates the value at the last moment; f t — represents the forget gate function; σ—represents the Sigmoid function; f t —Indicates that the output state is 0 to 1; a and b are coefficients; ② Input gate function: The input gate determines whether new information is placed in the cell state. It first passes through a tanh layer to obtain the representation of the current information. It also passes through a sigmoid layer to calculate which of the new information is important and which is not important, and then multiplies it with the output of the tanh layer and adds it to the cell state. This process, in short, is to weight the current new information into the cell state; After these operations, we can get the cell state at the current moment. The mathematical model in the input gate function and activation function is: i t =σ(W i ×[h t-1 ,x t ]+b i ) (30) Where: h t-1 —Indicates the state at the last moment; x t —Indicates the value at the last moment; σ—represents the forget gate function; tanh — represents the activation function; —Indicates the cell state at the current moment; ③ Output gate function: The output gate determines the output state at this moment. It first converts the cell state at the current moment After a tanh layer excitation, and then by h t-1 and x t After a Sigmoid layer, the updated weights are obtained, and the two are multiplied; the result is the current cell state weighted by tanh excitation to obtain the current state; Where: C t —Indicates that the current cell state is weighted after tanh excitation to obtain the current state; f t — indicates the result of the forget gate function; C t-1 — represents the state of the cell at the previous time after the weighted state after tanh excitation; —Indicates the cell state at the current moment.

Images