Multi-modal data driven urban solid waste incineration controlled object modeling method

Abstract

The invention provides a multi-modal data driven urban solid waste incineration controlled object modeling method. The method relates to the technical field of the MSWI process and comprises the following steps: acquiring a global quantized value of an initial flame combustion state; acquiring virtual data under multiple working conditions; the initial flame combustion state global quantized value is input into the improved DFR combustion state quantized model in combination with real process data, and a final flame combustion state global quantized value is obtained; acquiring real operation data and expanding the real process data by using a VSG technology to obtain expanded data; mixing the virtual data and the extended data under multiple working conditions to obtain multi-working-condition time sequence data under multiple scenes; and according to the multi-working-condition time sequence data under the multiple scenes and the final flame combustion state global quantized value, a multiple-input multiple-output controlled object model is constructed, and the multiple-input multiple-output controlled object model comprises stacked four layers of LSTMs. According to the invention, the problem of low modeling accuracy caused by small volume of the modeling sample of the controlled object in the MSWI process in the prior art is solved.

Description

Technical Field

[0001] The invention relates to the field of MSWI process technology, and in particular to a multi-modal data driven urban solid waste incineration controlled object modeling method. Background Art

[0002] The mechanism of the MSWI process is complex and unclear, and its controlled object model needs to be constructed based on multimodal data consisting of process data and flame images. Although some ACC systems are equipped with consumable optical devices that can detect the length of the flame combustion line in real time, this detection value cannot globally characterize the flame combustion state. Excellent experts in our field can achieve precise flame combustion control in some cases by visually perceiving flames. How to establish a complete database for quantifying the flame combustion state and achieve local and global quantification based on the unevenly distributed or even missing multi-condition flame images existing in industrial sites, combined with the furnace structure and expert knowledge, is still an open problem to be solved. Although the historical data of the actual MSWI process is large in volume, the conditions that can be characterized are difficult to describe and unevenly distributed, which leads to poor data availability. Therefore, the modeling of the controlled object still faces the problem of small samples, which needs to be solved by using relevant technologies. At the same time, how to match process data and image data under multiple conditions, how to use numerical simulation software to obtain more comprehensive virtual (mechanism) data than on-site conditions, and how to fuse the above data to build a controlled object model need to be solved.

[0003] The flame combustion state in the incinerator is the primary object of combustion process control. However, the domestic use of video acquisition to the control center in the form of expert visual perception makes it difficult to use it as a controlled variable and support modeling. This requires the completion of the flame combustion image library and the quantitative modeling of the combustion state to supplement the missing controlled variable detection values. In addition to the flame combustion state, the controlled variables of the MSWI process also include furnace temperature, oxygen content and steam flow. Although these controlled variables and their corresponding operating variables can be collected in real time, the operating conditions they represent are difficult to describe and unevenly distributed. This requires numerical simulation multi-condition analysis and virtual data generation to obtain virtual data under multiple conditions. The MSWI process contains real images and process data as well as generated virtual data. It is necessary to conduct a study on the modeling of the controlled object by integrating multi-source data hybrid drive to build a multi-input and multi-output controlled object model that can be explained and conforms to the process mechanism. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a multimodal data driven urban solid waste incineration controlled object modeling method. The present invention solves the problem in the prior art that the controlled object modeling sample volume of the MSWI process is small and the available data is poor, resulting in low modeling accuracy.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A multimodal data driven urban solid waste incineration controlled object modeling method, comprising:

[0007] Obtain the global quantized value of the initial flame combustion state;

[0008] Obtain virtual data under multiple working conditions;

[0009] In combination with real process data, the initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model to obtain the final flame combustion state global quantization value;

[0010] Acquire real operation data and use VSG technology to expand real process data to obtain expanded data;

[0011] Mixing the virtual data under the multiple working conditions and the expanded data to obtain multi-working condition time series data under multiple scenarios;

[0012] A multi-input multi-output controlled object model is constructed according to the multi-operating condition time series data under the multiple scenarios and the final global quantization value of the flame combustion state, wherein the multi-input multi-output controlled object model includes a stacked four-layer LSTM.

[0013] Preferably, the obtaining of virtual data under multiple working conditions includes:

[0014] Conduct MSW sampling and analyze MSW samples to obtain process input factors;

[0015] According to the multi-condition orthogonal test design and the process input factors, a multi-factor and multi-level orthogonal experimental table is obtained;

[0016] Determine the boundary conditions for numerical simulation;

[0017] According to the numerical simulation boundary conditions, FLIC was used to simulate the solid phase combustion of MSW on the grate to obtain the solid phase combustion products.

[0018] FLUENT was used to simulate the gas phase combustion of MSW in the furnace and obtain the gas phase combustion products;

[0019] A solid-gas phase coupled combustion simulation is performed according to the solid-phase combustion products and the gas-phase combustion products, and based on the input conditions of the solid-gas phase coupled combustion simulation, ASPEN Plus is used to simulate the material energy balance of the whole process and virtual data under multiple working conditions are obtained according to a multi-factor and multi-level orthogonal experimental table.

[0020] Preferably, the solid phase combustion products include:

[0021] Flue gas temperature and flue gas composition.

[0022] Preferably, the gas phase combustion products include:

[0023] Gas temperature field, velocity field and concentration field.

[0024] Preferably, the initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model in combination with the real process data to obtain the final flame combustion state global quantization value, including:

[0025] Acquire real process data and fuse multiple frames of flame video according to the time scale of the real process data to obtain a global quantization value of the initial flame combustion state;

[0026] The initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model and aligned with the real process data to perform data combination to obtain the final flame combustion state global quantization value.

[0027] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0028] The present invention provides a multimodal data driven urban solid waste incineration controlled object modeling method, comprising:

[0029] Obtain the global quantization value of the initial flame combustion state; obtain virtual data under multiple working conditions; combine the real process data, input the global quantization value of the initial flame combustion state into the improved DFR combustion state quantization model to obtain the final global quantization value of the flame combustion state; obtain the real operation data and use the VSG technology to expand the real process data to obtain the expanded data; mix the virtual data under the multiple working conditions and the expanded data to obtain the multi-condition time series data under multiple scenes; construct a multi-input multi-output controlled object model according to the multi-condition time series data under the multiple scenes and the final global quantization value of the flame combustion state, wherein the multi-input multi-output controlled object model includes a stacked four-layer LSTM. The present invention uses the VSG technology to expand the real data and mix it with the multi-condition virtual data to form the time series data of multiple scenes, thereby increasing the data diversity during model training, helping to capture the complex characteristics of the incineration process, increasing the volume of modeling samples, and enhancing the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1A flow chart of a multimodal data driven urban solid waste incineration controlled object modeling method provided by the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, the present invention provides a multimodal data driven urban solid waste incineration controlled object modeling method, comprising:

[0035] Step 100: Obtaining a global quantized value of an initial flame combustion state;

[0036] Step 200: Acquire virtual data under multiple working conditions;

[0037] Step 300: In combination with real process data, the initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model to obtain the final flame combustion state global quantization value;

[0038] Step 400: Acquire real operation data and use VSG technology to expand the real process data to obtain expanded data;

[0039] Step 500: Mixing the virtual data under the multiple working conditions and the expanded data to obtain multi-working condition time series data under multiple scenarios;

[0040] Step 600: construct a multi-input multi-output controlled object model according to the multi-operating condition time series data under the multiple scenarios and the final global quantization value of the flame combustion state, wherein the multi-input multi-output controlled object model includes a stacked four-layer LSTM.

[0041] Specifically, according to the expert experience of engineers in the field, the geometric segmentation theory is used to qualitatively analyze the combustion characteristics (flame height, brightness and length) presented in the flame image. The quantitative value obtained by the first module is one of the outputs of the modeling of the third module. Then, the typical images are manually annotated to obtain a limited small sample data set. The flame image data is automatically calibrated using the visual Transformer. GAN is used in combination with the flame imaging mechanism to obtain atypical images. A complete flame combustion image quantization database is established. The calibrated complete flame combustion images are matched with the operating variables of the same time scale to obtain a training data set. A multi-input and multi-output deep forest regression (DFR) model that mixes multi-layer stacking learning and selective attention mechanism is constructed. Feature fusion is used in the output layer to realize the mapping between the local quantitative value and the comprehensive quantitative value of the flame combustion state. Based on the component attributes of MSW and M The combustion characteristics of the SWI process are used to analyze the key factors, and multi-condition operation scenarios are designed according to the knowledge of domain experts. On this basis, a multi-factor and multi-level orthogonal experimental design is used to obtain an orthogonal test table. Considering the physical and chemical properties of solid-phase combustion, gas-phase fluid dynamics and full-process mass-energy balance reactions contained in the MSWI process, a full-process simulation model is constructed using a multi-numerical simulation software coupling strategy, and experiments are carried out based on the orthogonal test table to obtain virtual (mechanism) data under multiple conditions; for the MSWI multi-condition virtual (mechanism) data obtained by multi-software coupling and the real data of benchmark conditions selected from industrial sites, a data fusion mechanism based on domain expert knowledge and virtual sample generation (VSG) technology are used to obtain a multi-scenario sample library for modeling, a serial integrated model structure is designed according to the process mechanism, and a multi-input and multi-output timing characteristic controlled object model is constructed based on the obtained sample library.

[0042] Furthermore, expert knowledge is used to calibrate flame combustion images. Based on domain expert knowledge and operational experience, flame combustion images are quantized based on geometric theory. The flame combustion state is quantized into three local quantization values ​​of height, brightness, and length and one global quantization value of the flame combustion state. Then, the combustion images are manually labeled according to the four quantization targets to obtain a typical flame combustion image quantization data set. The flame combustion image automatic calibration model takes the obtained typical image data set as input, pre-processes the image through image linear block segmentation and linear mapping, and then trains and obtains a flame combustion image automatic calibration model that combines a Transformer encoder and a multi-layer perceptron. Unlabeled flame combustion images Image automatic calibration, frame rate extraction of flame video to obtain the flame combustion image data to be labeled, input into the automatic calibration model, and obtain the calibrated flame combustion image quantization data set; mechanism knowledge and generative adversarial network (GAN) pseudo-labeled atypical flame image quantization set, combined with the mapping mechanism of three-dimensional space imaging in the furnace, use the typical image data set to obtain pseudo-labeled atypical images, combined with unlabeled flame images, use the improved cycle consistency GAN and generated image evaluation criteria to obtain the atypical image data set; construct a complete flame combustion state quantization database, merge the typical image data set, the automatic calibration image data set and the atypical image data set to obtain a complete flame combustion state quantization value database.

[0043] The data set is constructed with the quantized value of the flame combustion state as the dependent variable (true value) and the wind cloth operation variable as the independent variable; the DFR model structure is determined, including the type, number and stacking number of forest algorithms in each layer; the attention mechanism DFR model is selected to construct a multi-input and multi-output improved DFR model with three local quantized values ​​of flame height, brightness and length as output, and the selective attention mechanism is used between layers to improve the expressiveness of features; the flame height, brightness and length values ​​are fused, and a linear fusion strategy is adopted based on the three local quantized values ​​of height, brightness and length, and the final global quantized value of the flame combustion state is obtained in combination with the operation process state information.

[0044] Furthermore, the obtaining of virtual data under multiple working conditions includes:

[0045] Conduct MSW sampling and analyze MSW samples to obtain process input factors;

[0046] According to the multi-condition orthogonal test design and the process input factors, a multi-factor and multi-level orthogonal experimental table is obtained;

[0047] Determine the boundary conditions for numerical simulation;

[0048] According to the numerical simulation boundary conditions, FLIC was used to simulate the solid phase combustion of MSW on the grate to obtain the solid phase combustion products.

[0049] FLUENT was used to simulate the gas phase combustion of MSW in the furnace and obtain the gas phase combustion products;

[0050] A solid-gas phase coupled combustion simulation is performed according to the solid-phase combustion products and the gas-phase combustion products, and based on the input conditions of the solid-gas phase coupled combustion simulation, ASPENPlus is used to simulate the material energy balance of the whole process and virtual data under multiple working conditions are obtained according to a multi-factor and multi-level orthogonal experimental table.

[0051] Specifically, for MSW sampling, the fermented and dehydrated MSW is manually sampled based on the four-partition method proposed in the "Sampling and Analysis Methods for Municipal Waste" to obtain MSW samples for analysis; for MSWI process input factor analysis, the obtained MSW samples are tested offline to analyze the MSW calorific value (low calorific value), industrial analysis (moisture, ash, volatile matter, fixed carbon) and elemental analysis (C, H, O, N, S) data and the feasible intervals of MSW composition factors; for MSWI process operation factor analysis, based on the three states of low load, full load and overload, the possible operating conditions in the operation of the MSWI process are analyzed based on the knowledge of domain experts, and then the feasible intervals of the operation factors are obtained; for multi-condition orthogonal experiment design, the MSW composition factors and operation factors are used as factors of the multi-condition orthogonal experiment, the number of levels of the orthogonal experiment is calculated based on the feasible interval, and then the statistical analysis software is used to obtain the multi-factor and multi-level orthogonal experiment table to provide support for the full process numerical simulation.

[0052] More specifically, determine the boundary conditions of numerical simulation, perform numerical simulation on solid phase combustion, gas phase combustion and material energy balance according to the process design of the actual MSWI process, and determine the input boundary conditions of the orthogonal experiment based on the operation process; for MSW solid phase combustion simulation, use FLIC to simulate the solid phase combustion of MSW on the grate, obtain the flue gas data generated by solid phase combustion, the boundary conditions of gas phase combustion and the flue gas temperature and flue gas components required for material balance simulation; for MSW gas phase combustion simulation, use FLUENT to simulate the gas phase combustion of MSW in the furnace, obtain the furnace radiation data and material balance required for solid phase MSW combustion, and the gas temperature field, velocity field and concentration field required for simulation; for solid-gas phase coupled combustion simulation, interact with the solid-gas phase combustion model until the boundary conditions and radiation data converge, and obtain the flue gas temperature, flue gas components, gas temperature field, velocity field and concentration field; for the whole process material balance simulation, use ASPEN based on the input conditions of solid-gas phase coupled combustion simulation. Plus simulates the material and energy balance of the whole process, obtains the numerical simulation model of the whole process and generates furnace temperature, steam flow and oxygen content as well as pollution emission data; multi-condition orthogonal experimental research, carries out numerical simulation of the whole process according to the obtained multi-factor and multi-level orthogonal experimental table, and then generates virtual data under multiple conditions.

[0053] Furthermore, the initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model in combination with the real process data to obtain the final flame combustion state global quantization value, including:

[0054] Acquire real process data and fuse multiple frames of flame video according to the time scale of the real process data to obtain a global quantization value of the initial flame combustion state;

[0055] The initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model and aligned with the real process data to perform data combination to obtain the final flame combustion state global quantization value.

[0056] Specifically, the flame combustion image state quantization value calibration takes the time scale of the real process data as the benchmark, fuses multiple frames of the flame video, obtains the flame combustion image that needs to be quantized, inputs the improved DFR combustion state quantization model, and obtains the global quantization value of the flame combustion state; the flame combustion image data is aligned with the real process data, and the obtained global quantization value of the flame combustion state is combined with the current real process data to obtain the operating condition data set containing the control variables and the controlled variables at the current moment; the real data distribution boundary filling based on virtual sample generation (VSG) is used to store the obtained real operating condition data. In order to solve the problem of uneven distribution, the virtual (mechanism) data generated by numerical simulation under similar operating conditions is combined with VSG technology to expand the real operating data, and obtain the complete distribution data under the real benchmark conditions; multi-condition virtual data preprocessing, anomaly detection and abnormal sample deletion processing are performed on the obtained multi-condition virtual data, and then the multi-condition data is segmented to obtain multiple single-condition data. Combined with the flame combustion image complete database constructed in the previous article, the same strategy as above is used to expand the local and global quantization values ​​of flame combustion, and the complete distribution data under multiple virtual single conditions is obtained; multi-condition data scenario Classification, in order to restore the real MSWI process operation state, based on the domain expert knowledge and MSWI process operation characteristics, according to the multi-scenario requirements and data timing, the multi-condition virtual data and the benchmark real process data are mixed and recombined to obtain multi-condition timing data under multiple scenarios; the multi-input and multi-output controlled object model structure design, according to the MSWI process flow design serial integrated output model structure, the order is: flame combustion state → furnace temperature → steam flow → oxygen content; in the flame combustion state model part, the air distribution and material distribution control amount is used as the input, the first layer of LSTM is constructed, and the flame combustion state quantization value output is obtained Output; For the furnace temperature model part, the air distribution and material distribution control amount and the quantized value of the flame combustion state are used as input to construct the second layer of LSTM, and the furnace temperature output is obtained; For the steam flow model part, the air distribution and material distribution control amount, the flame combustion state and the furnace temperature are used as input to construct the third layer of LSTM, and the steam flow output is obtained; For the oxygen content model part, the air distribution and material distribution control amount, the flame combustion state, the furnace temperature and the steam flow are used as input to construct the fourth layer of LSTM, and the oxygen content output is obtained; For the multi-input and multi-output controlled object model, a controlled object model with timing characteristics is obtained by serially stacking four layers of LSTM.

[0057] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0058] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A multimodal data driven modeling method for a controlled object of municipal solid waste incineration, characterized in that: include: Obtain the global quantized value of the initial flame combustion state; Obtain virtual data under multiple working conditions; In combination with real process data, the initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model to obtain the final flame combustion state global quantization value; Acquire real operation data and use VSG technology to expand real process data to obtain expanded data; Mixing the virtual data under the multiple working conditions and the expanded data to obtain multi-working condition time series data under multiple scenarios; A multi-input multi-output controlled object model is constructed according to the multi-operating condition time series data under the multi-scenario and the final global quantization value of the flame combustion state, wherein the multi-input multi-output controlled object model includes a stacked four-layer LSTM.

2. According to claim 1, a multimodal data driven urban solid waste incineration controlled object modeling method is characterized in that: The obtaining of virtual data under multiple working conditions includes: Conduct MSW sampling and analyze MSW samples to obtain process input factors; According to the multi-condition orthogonal test design and the process input factors, a multi-factor and multi-level orthogonal experimental table is obtained; Determine the boundary conditions for numerical simulation; According to the numerical simulation boundary conditions, FLIC was used to simulate the solid phase combustion of MSW on the grate to obtain the solid phase combustion products. FLUENT was used to simulate the gas phase combustion of MSW in the furnace and obtain the gas phase combustion products; A solid-gas phase coupled combustion simulation is performed according to the solid-phase combustion products and the gas-phase combustion products, and based on the input conditions of the solid-gas phase coupled combustion simulation, ASPEN Plus is used to simulate the material energy balance of the whole process and virtual data under multiple working conditions are obtained according to a multi-factor and multi-level orthogonal experimental table.

3. The multimodal data driven urban solid waste incineration controlled object modeling method according to claim 2 is characterized in that: The solid phase combustion products include: Flue gas temperature and flue gas composition.

4. The multimodal data driven urban solid waste incineration controlled object modeling method according to claim 2 is characterized in that: The gas phase combustion products include: Gas temperature field, velocity field and concentration field.

5. The multimodal data driven urban solid waste incineration controlled object modeling method according to claim 2 is characterized in that: The initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model in combination with the real process data to obtain the final flame combustion state global quantization value, including: Acquire real process data and fuse multiple frames of flame video according to the time scale of the real process data to obtain a global quantization value of the initial flame combustion state; The initial flame combustion state global quantization value is input into the improved DFR combustion state quantization model and aligned with the real process data to perform data combination to obtain the final flame combustion state global quantization value.

Images