An evaporation crystallization process optimization method and system based on multi-source data fusion

Abstract

The present application belongs to the field of industrial process control, and relates to an evaporation crystallization process optimization method and system based on multi-source data fusion. The method comprises the following steps: collecting multi-source time series data of the evaporation crystallization system and constructing an evaporation load index; training a working condition decoupling model based on historical health data, and using the model to predict the theoretical health heat exchange coefficient in real time; calculating the residual error between the theoretical health heat exchange coefficient and the real-time observed heat exchange coefficient, extracting the pure scaling loss signal, and predicting the future trend of the signal; constructing a total average loss rate function containing cumulative operation loss and shutdown opportunity loss, and solving the minimum value of the function to determine the optimal cleaning time point. The present application solves the problem that the prior art cannot distinguish between process loss and physical loss, leading to inaccurate cleaning decision, and significantly improves the operation efficiency and economy of the system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial process control, and particularly relates to an evaporation crystallization process optimization method and system based on multi-source data fusion. BACKGROUND

[0002] As a kind of efficient and energy-saving evaporation equipment, MVR evaporator is widely used in the treatment of high-salt and high-COD industrial wastewater such as chemical industry and pesticide industry. Its working principle is to use a steam compressor to adiabatically compress the secondary steam, thereby improving the steam temperature and pressure, and then using the steam to heat the liquid, which greatly reduces the energy consumption.

[0003] However, in the evaporation crystallization process, a core problem that has long plagued production is the scaling and clogging of the heat exchange pipeline. The salt and organic matter contained in the industrial wastewater will continue to precipitate and form a hard scale layer on the inner wall of the heat exchange pipe during evaporation and concentration. This scaling phenomenon is not only an appearance problem, but also directly causes two serious production consequences: first, the energy consumption soars, the scale layer is a poor conductor of heat, which leads to a sharp decrease in the total heat exchange coefficient (K value), in order to maintain the rated evaporation capacity, the steam compressor, the core component of the MVR system, must consume more electricity to improve the temperature and pressure, resulting in a significant reduction in system energy efficiency; second, unplanned shutdown is caused, when the scaling is serious enough to block the pipeline, the entire system will be forced to shut down for non-planned maintenance, which requires time and effort for chemical or physical cleaning, seriously affecting the continuity of production.

[0004] At present, the decision-making method of factories for when the equipment needs to be cleaned is generally rough, mainly relying on two ways. The first is passive maintenance, that is, waiting until the equipment operating parameters (such as evaporation capacity) have decreased significantly or the blockage alarm is triggered before stopping for cleaning. This method has actually caused a lot of energy waste and potential equipment damage. The second is fixed-period maintenance, such as cleaning once every 30 days. This one-size-fits-all approach cannot adapt to complex and variable actual working conditions. When treating high-concentration wastewater, the equipment may be severely scaled in only 15 days; when treating low-concentration wastewater, the equipment may not need to be cleaned even after 60 days of operation. This approach often leads to unnecessary downtime and loss of production hours.

[0005] To overcome the limitations of traditional maintenance methods, some technicians have attempted to use time-series algorithms (such as ARIMA models and basic neural networks) to predict the decreasing trend of the heat transfer coefficient (K value) for predictive maintenance. However, these existing technologies have a key technical flaw: they cannot effectively decouple "reversible process efficiency loss" from "irreversible physical scaling loss." The real-time K value of the evaporator is affected not only by the slowly accumulating physical scaling but also by drastic fluctuations in operating conditions such as feed flow rate, feed concentration, and steam pressure. For example, a sudden increase in feed concentration can cause a temporary decrease in the K value, but this is not physical scaling. Traditional models conflate the two types of loss, misjudging the decrease in K value caused by operating condition fluctuations as physical scaling accumulation, resulting in a biased assessment of the actual scaling degree and thus failing to provide an accurate optimal cleaning time. Summary of the Invention

[0006] Therefore, the purpose of this invention is to propose an optimization method and system for the evaporation and crystallization process based on multi-source data fusion, in order to solve the technical problem in the prior art that the process loss and physical scaling loss cannot be decoupled, resulting in inaccurate decision-making on the timing of cleaning.

[0007] To address the above problems, the technical solution proposed in this invention for optimizing the evaporation and crystallization process based on multi-source data fusion is as follows:

[0008] The optimization method for the evaporation and crystallization process based on multi-source data fusion includes the following steps:

[0009] Collect multi-source time-series data of the evaporation crystallization system and construct an evaporation load index to characterize the comprehensive load under real-time operating conditions; the multi-source time-series data includes at least the feed flow rate, feed concentration, relevant parameters of the steam compressor, and real-time observed heat transfer coefficient;

[0010] Based on the historical health data of the evaporation and crystallization system, a decoupling model for operating conditions is trained; using the decoupling model, the theoretical healthy heat transfer coefficient that the system should have when it is in a healthy state under the current operating conditions is predicted according to the real-time collected operating condition variables.

[0011] By calculating the residual between the theoretical healthy heat transfer coefficient and the real-time observed heat transfer coefficient, the pure scaling loss signal is extracted, and based on the historical data of the pure scaling loss signal, the scaling loss trend in the future time period is predicted.

[0012] A total average loss rate function is constructed to characterize the balance between system operating energy consumption loss and downtime opportunity loss. The minimum value of the total average loss rate function is then calculated based on the scaling loss trend to determine the optimal cleaning time point for the evaporation crystallization system, thereby achieving dynamic optimization of the evaporation crystallization process.

[0013] Furthermore, the formula for calculating the evaporation load index is as follows:

[0014] ;in, for Evaporation load index at time of day and They are respectively The feed flow rate and feed concentration at any given time, and These are the historical steady-state average flow rate and the historical steady-state average concentration, respectively. and These are the sensitivity weighting factors for flow rate and concentration, respectively. It is a natural exponential function.

[0015] Furthermore, the operating condition decoupling model is a gated cyclic unit network model, and the historical health data is the multi-source time-series data within a preset time period after each chemical cleaning of the evaporation crystallization system.

[0016] Furthermore, the formula for calculating the pure scaling loss signal is as follows:

[0017] ;in, for The signal of constant cleanliness loss due to scaling. The theoretical healthy heat transfer coefficient output by the decoupled model under the stated operating conditions. The real-time observed heat transfer coefficient is collected in real time.

[0018] Furthermore, the prediction of scale loss trends over future time periods includes:

[0019] The pure scaling loss signal was fitted using a multinomial regression model. Historical data, and extrapolated to obtain future... Predicted scaling loss at each moment .

[0020] Furthermore, the total average loss rate function is:

[0021] ;in, The exact timeframe for future cleaning is yet to be determined. For the current moment, From arrive The cumulative operational losses due to scale buildup over time To avoid downtime and potential losses, To fix the cleaning time, For runtime, This represents the average power under healthy conditions.

[0022] Further, the additional power consumption in the cumulative operation loss is , which is calculated by historical data fitting or thermodynamic formula, and represents the scaling loss , the steam compressor compared to the average power in the healthy state The additional power consumption required.

[0023] Further, the optimal cleaning time point is the minimum value of the total average loss rate function value.

[0024] Further, the steam compressor related parameters at least include the steam compressor outlet pressure and the steam compressor operating current.

[0025] The technical scheme of the evaporation crystallization process optimization system based on multi-source data fusion provided by the present application is:

[0026] The evaporation crystallization process optimization system based on multi-source data fusion comprises the following modules:

[0027] The data acquisition module is used to acquire multi-source time series data of the evaporation crystallization system, and to construct an evaporation load index for representing the comprehensive load of real-time working conditions; the multi-source time series data at least includes feed flow, feed concentration, steam compressor related parameters and real-time observed heat transfer coefficient;

[0028] The model training and prediction module is used to train a working condition decoupling model based on the historical health data of the evaporation crystallization system; and to predict the theoretical health heat transfer coefficient that the system should have in the healthy state under the current working condition according to the real-time acquired working condition variables by using the working condition decoupling model;

[0029] The scaling extraction and prediction module is used to extract a pure scaling loss signal by calculating the residual between the theoretical health heat transfer coefficient and the real-time observed heat transfer coefficient, and to predict the scaling loss trend in the future period based on the historical data of the pure scaling loss signal;

[0030] The optimization decision module is used to construct a total average loss rate function representing the balance relationship between the system operation energy consumption loss and the shutdown opportunity loss, and to determine the optimal cleaning time point of the evaporation crystallization system according to the minimum value of the total average loss rate function solved based on the scaling loss trend, so as to realize the dynamic optimization of the evaporation crystallization process.

[0031] Further, the calculation formula of the evaporation load index is:

[0032] ; wherein, is the evaporation load index at the moment , and and are respectively the feed flow rate and the feed concentration at time t, and are the historical steady-state average flow rate and the historical steady-state average concentration, respectively, and are the sensitivity weight factors for flow rate and concentration, respectively, is a natural exponential function.

[0033] Further, the working condition decoupling model is a gated recurrent unit network model, and the historical health data is the multi-source time series data in a preset time period after each chemical cleaning of the evaporative crystallization system is completed.

[0034] Further, the calculation formula of the pure fouling loss signal is:

[0035] ; wherein, is the pure fouling loss signal at time t, is the theoretical health heat exchange coefficient output by the working condition decoupling model, is the real-time observed heat exchange coefficient collected in real time.

[0036] Further, the prediction of the fouling loss trend in the future time period comprises:

[0037] a polynomial regression model is used to fit historical data of the pure fouling loss signal , and the predicted fouling loss at the next time points is extrapolated.

[0038] Further, the total average loss rate function is:

[0039] ; wherein, is the future cleaning time point to be determined, is the current time, is the cumulative operating loss accumulated due to fouling from time to time , is the shutdown opportunity loss, is the fixed cleaning time, is the operating time, is the average power in the healthy state.

[0040] Further, the additional power consumption in the cumulative operating loss is , which is obtained by fitting historical data or a thermodynamic formula, and represents that when the fouling loss is , the steam compressor has a power consumption compared to the average power in the healthy state.The additional power required for consumption.

[0041] Further, the optimal cleaning time point is the minimum value of the total average loss rate function Value.

[0042] Further, the steam compressor related parameters at least include the steam compressor outlet pressure and the steam compressor operating current.

[0043] The beneficial effects of the present application are: the present application constructs the evaporation load index by fusing the nonlinear relationship of feed flow and concentration, accurately characterizes the real-time working condition comprehensive load, and provides high-quality input basis for subsequent working condition decoupling; relying on the GRU network model to train the historical data of the health period after cleaning of the evaporation crystallization system, fully exerting the advantages of capturing the nonlinear dependence and time lag effect of time series data, realizing the accurate prediction of the theoretical health heat exchange coefficient under the current working condition, and then comparing the theoretical health heat exchange coefficient with the real-time observed heat exchange coefficient through the residual method, successfully stripping the reversible process loss interference such as feed fluctuation and steam pressure change, and extracting the pure scaling loss signal reflecting only the irreversible physical scaling accumulation, completely solving the core pain point that the existing technology cannot distinguish between the two types of loss and leads to scaling evaluation deviation.

[0044] The present application helps the polynomial regression model to fit and extrapolate the historical data of the pure scaling loss signal, accurately predicts the future scaling development trend, and constructs the total average loss rate function by fusing the cumulative operating energy loss and downtime opportunity loss, finds the optimal cleaning time point by solving the minimum value of the function, balances the energy waste and downtime cost, not only completely replaces the excessive energy consumption, equipment additional damage and unplanned downtime risk caused by passive maintenance, but also avoids unnecessary downtime loss or excessive scaling accumulation caused by fixed cycle maintenance, significantly improves the heat exchange efficiency and operation stability of the MVR evaporation crystallization system, greatly reduces the additional power consumption of the steam compressor and the equipment maintenance cost, effectively improves the equipment utilization and production continuity, and at the same time, the data-driven dynamic optimization logic adapts to the working condition fluctuation of different concentrations and flow rates, has strong industrial applicability and flexibility, and provides a more intelligent, efficient and economic technical solution for the evaporation crystallization process of high-salt and high-COD wastewater treatment in the fields of chemical industry and pesticide industry. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The step flow chart of the evaporation crystallization process optimization method based on multi-source data fusion of the present application;

[0046] Figure 2 The decoupling analysis diagram of the heat exchange coefficient K value;

[0047] Figure 3 The pure scaling loss signal diagram;

[0048] Figure 4 A schematic diagram for optimal cleaning cycle decision. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be clearly and completely described in combination with the drawings in the embodiments of the present application.

[0050] The specific embodiments of the evaporation crystallization process optimization method based on multi-source data fusion provided in the present application are as follows:

[0051] As shown in the figure, the evaporation crystallization process optimization method based on multi-source data fusion is run on an MVR evaporation crystallization system, and specifically includes the following steps: Figure 1 S1, collecting multi-source time series data of the evaporation crystallization system, and constructing an evaporation load index for representing a real-time working condition comprehensive load; the multi-source time series data at least includes feed flow, feed concentration, steam compressor related parameters and real-time observed heat exchange coefficient.

[0052] First, sensors are deployed at key nodes of the MVR evaporation crystallization system to synchronously collect multi-source time series data streams in real time. These multi-source time series data are divided into working condition variables and target variables. The working condition variables are variable inputs describing the process state, at least including: feed flow, feed concentration, steam compressor outlet pressure

[0053] , steam compressor operating current . The target variable is the total heat exchange coefficient , which is calculated in real time through a thermodynamic formula , and the thermodynamic formula can be transformed as follows: ; wherein, is the effective heat transfer area of the heat exchanger, and is a fixed design parameter of the heat exchanger; is the logarithmic mean temperature difference, which is calculated by using the collected steam temperature (hot side) and heat exchanger inlet and outlet temperature (cold side); is the heat exchange rate / heat load, which is calculated by using the circulating flow, fluid specific heat capacity and heat exchanger inlet and outlet temperature difference, etc. In the thermodynamic formula, the calculation of the remaining parameters except the parameter K is the prior art, which will not be described in detail here. It should be noted that, is a signal mixed with "working condition fluctuation" and "physical fouling" effects.

[0054] ​In data acquisition, directly inputting feed flow rate and feed concentration as independent features into the subsequent model ignores their physical multiplicative effect. For example, the combined pressure of high flow rate and high concentration is much greater than that of low flow rate and high concentration. Therefore, this step constructs a fusion feature, namely the evaporation load index, to characterize this comprehensive evaporation pressure. The formula for calculating the evaporation load index is:

[0055] ;in, for The evaporation load index at any given time is a dimensionless scalar. yes The feed flow rate at any given time. yes The feed concentration at any given time; and The historical steady-state average flow rate and historical steady-state average concentration are used as the normalization benchmarks. and The sensitivity weighting factors for flow rate and concentration are determined based on expert experience or small-scale regression tests. It is a natural exponential function.

[0056] This formula doesn't simply add the flow rate and concentration parameters together; instead, it uses multiplication and exponential methods to construct a mathematical model that conforms to the laws of physics, thus more realistically reflecting the pressure experienced by the MVR system. The flow rate is expressed through a linear proportional term. The influence on evaporation load follows the physical law that doubling the flow rate, while keeping the concentration constant, roughly doubles the evaporation load. Concentration is affected by the exponential term. The evaporation load is affected because an increase in concentration leads to a non-linear rise in the solution's boiling point, reducing the effective temperature difference required for evaporation and exponentially increasing the difficulty of evaporation. Multiplying these two factors, we can recreate the physical reality that "a small increase in concentration at high flow rates" results in a greater system load than "at low flow rates."

[0057] For example, assume historical steady-state average flow Historical steady-state average concentration Traffic sensitivity weighting factor Concentration sensitivity weighting factor .exist At any time, collected , , Substituting the above data into the formula for calculating the evaporation load index, we can obtain... =0.5639.

[0058] Thus, by constructing the evaporation load index, multiple operating condition variables are fused into a nonlinear feature that can highly represent the physical load of the system, providing high-quality input for the subsequent precise decoupling of the model.

[0059] S2, based on the historical health data of the evaporation crystallization system, training an operating condition decoupling model; using the operating condition decoupling model, predicting the theoretical health heat exchange coefficient that the system should have in a healthy state according to the real-time collected operating condition variables.

[0060] The purpose of this step is to train a model to learn the reversible relationship between "operating condition" and "K value", that is, to build a digital twin health model that does not consider fouling.

[0061] First, the training data needs to be screened. From the historical database of the MVR evaporation crystallization system, all data in the healthy state are screened. The healthy state is defined as a pre-set time period after each chemical cleaning, for example, the first 48 hours. Within these 48 hours, the physical fouling At this time, the fluctuation of K value is only caused by the operating condition variables.

[0062] Then, the training set is constructed. Assuming that 100 cleaning cycles are collected, and 48 hours of data are taken each time, a large health data set is spliced.

[0063] Next, the model is trained. A gated recurrent unit (GRU) network model is constructed, denoted as GRU-P. GRU is an advanced recurrent neural network (RNN) that contains a reset gate and an update gate, making it good at capturing nonlinear dependencies and time lag effects in time series data. For example, the current high concentration may not fully reflect on the K value for several minutes.

[0064] The input of the GRU-P model is a sequence of N time steps of operating condition variables, i.e., for the current time , the model receives all data points from time to time , which contains the following variables: evaporation load index from step S1, steam compressor outlet pressure and compressor operating current . For example, when N=10, the input is a sequence of operating condition data of the past 10 consecutive time steps. The output of the GRU-P model is the health K value at time , i.e. . By training on the health data set, the GRU-P model can accurately learn this complex reversible physical relationship.

[0065] In the MVR evaporation crystallization system enters the normal operation, that is, the non-training stage, the trained GRU-P model is used for real-time prediction. In time, the real-time working condition variable and the data of the previous N-1 time are input into the GRU-P model, and the GRU-P model outputs a predicted value . The physical meaning of is: assuming that the pipeline is completely clean, under the working condition at the current time , the theoretical value that the heat exchange coefficient should reach, that is, the theoretical healthy heat exchange coefficient output by the GRU-P model, is a healthy baseline of digital twinning.

[0066] In this way, by training the GRU-P model on pure healthy data, a digital twin that can reflect the theoretical healthy heat exchange coefficient in real time under the current working condition is obtained, which is the core benchmark for realizing subsequent working condition decoupling.

[0067] S3, by calculating the residual between the theoretical healthy heat exchange coefficient and the real-time observed heat exchange coefficient, a pure scaling loss signal is extracted, and based on the historical data of the pure scaling loss signal, a scaling loss trend in a future time period is predicted.

[0068] At time, step S2 obtains the theoretical healthy heat exchange coefficient , and step S1 collects the real-time observed heat exchange coefficient . This step calculates the residual between the two to strip the influence of working condition fluctuations and extract the pure scaling loss.

[0069] The pure scaling loss signal is calculated as follows: ; wherein, is the pure scaling loss signal at time, that is, the K value loss caused by irreversible physical scaling at time, with the unit of . is the theoretical healthy heat exchange coefficient predicted by the GRU-P model, that is, the theoretical K value; is the real-time observed heat exchange coefficient actually collected by the sensor, that is, the observed K value.

[0070] The logic of this residual decoupling is that when the working condition fluctuation causes to drop, the GRU-P model also receives the working condition fluctuation data, and its predicted will also drop, and remains unchanged; only when real scaling occurs, causing to be lower than hour, The value will only increase if... Therefore, It is a smooth, monotonically increasing curve that is almost unaffected by fluctuations in operating conditions; it only reflects the actual degree of physical scaling accumulation.

[0071] For example: in At any moment, assuming , ,at this time .exist At any given time, if operating conditions fluctuate, such as an increase in concentration, the model predicts... At the same time, the observed values ​​also decreased. ,at this time It can be observed that the scaling loss remained unchanged, and the interference from the operating conditions was successfully eliminated.

[0072] However, in At that moment, the working conditions were restored. ,but arrive A small amount of actual scaling occurred during this period, resulting in the observed values. ,at this time Scale loss increased by 5%.

[0073] Next, it is necessary to predict this scaling trend. Because... The curves are very smooth and exhibit strong trends, eliminating the need for complex GRUs. A simple multinomial regression or ARIMA model can be used to fit the curves. Historical data, such as the past 24 hours The value is extrapolated to obtain the future. Hourly predicted scaling loss ,in, From the current moment To the future The time.

[0074] At the same time, it is also necessary to establish scaling loss. With additional power consumption Relationship Model In this embodiment, ,in, The scale loss is The power of the steam compressor at that time can be determined according to... Estimate, The average power under healthy conditions; Indicates scaling loss as At that time, the steam compressor's average power compared to its healthy state The additional power required. This model uses thermodynamic formulas or historical data. and is fitted. is a non-linear increasing function of , i.e. the fouling is more serious, the additional power consumption grows faster.

[0075] Thus, by the GRU-P residual method, the pure fouling signal is successfully extracted, and its future trend and the corresponding energy loss model are obtained, providing the necessary input for the optimization decision of step S4.

[0076] S4, a total average loss rate function representing the balance between the energy loss and the downtime opportunity loss of the system is constructed, and the minimum value of the total average loss rate function is solved according to the fouling loss trend, to determine the optimal cleaning time point of the evaporative crystallization system, and realize the dynamic optimization of the evaporative crystallization process.

[0077] The goal of this step is to find an optimal cleaning cycle , i.e. the future time from now , so that the total average loss rate from to is the lowest.

[0078] In this embodiment, a total average loss rate function is constructed:

[0079] ; is the total average loss rate of cleaning at future time, in the denominator, is the future cleaning time point to be determined, is the current time, is the fixed cleaning time, is the running time, is the total cycle; the numerator contains two parts, the first part is , representing the cumulative running loss accumulated from to due to fouling, i.e. from to , according to the fouling trend and the energy loss model predicted in step S3, the total electric energy consumed due to fouling accumulation is calculated. is the downtime opportunity loss, which is the basic electric energy consumed by the system in a healthy state within the fixed cleaning time , which represents the equivalent energy loss caused by downtime; is the average power in a healthy state.

[0080] is a U-shaped curve, if is too short, i.e. frequent cleaning, is small, but is apportioned to a very short total period, resulting in is very high, in which case the opportunity loss of shutdown dominates. If is too long, i.e. delayed cleaning, is apportioned well, but will increase sharply, resulting in is also very high, in which case the cumulative running loss dominates.

[0081] The optimization decision of this step is to solve in real time (for example, every hour), i.e. to find the lowest point of the U-shaped curve. The value of at this lowest point is the optimal cleaning time point .

[0082] Exemplarily, assume that the current time is , the fixed cleaning time is , and the average power in the healthy state is . At this time, the opportunity loss of shutdown = 100 .

[0083] The system performs the following steps in real-time operation every hour:

[0084] Set the range of the variable , and make assumptions for different time points in the future, for example, calculate for different cases such as 24 hours, 100 hours, 200 hours, etc. in the future, and obtain , , by calculation. By comparison, is much lower than and . The system will continue to search and eventually find the lowest point of the U-shaped curve, for example = 100 hours, which is taken as .

[0085] The system finally displays to the operator in real time: "the current real fouling loss: ", "the predicted optimal cleaning time: hours later".

[0086] The effects of the present application can be understood in combination with the accompanying Figure 2 to Figure 4 .

[0087] Refer to Figure 2 , Figure 2is the decoupling analysis diagram of heat exchange coefficient K value. The X-axis is the running time, and the Y-axis is the heat exchange coefficient K value. The blue curve is the real-time observed K value , which fluctuates sharply and shows a downward trend as a whole, mixing "working condition noise" and "scaling trend". The green curve is the healthy K value predicted by the GRU-P model in step S2 , which also fluctuates sharply because it follows the changes in "working condition noise" in real time. In the chart, the blue curve is always systematically lower than the green curve, and the gap between the two curves becomes larger and larger over time, which is the result of scaling.

[0088] Referring to Figure 3 , Figure 3 is a pure scaling loss signal diagram, which shows how to clean the data by the algorithm to obtain a pure scaling signal. The X-axis is the running time, and the Y-axis is the scaling loss . The red curve is the calculation result of . Figure 3 All sharp fluctuations (common fluctuations of the blue and green curves) in are completely offset after subtraction. The red curve is a very smooth and monotonically increasing curve, which accurately evaluates the true accumulation of irreversible physical scaling. The purple dashed line is the future trend prediction of the red curve in step S3 .

[0089] Referring to Figure 4 , Figure 4 is an optimal cleaning cycle decision diagram. The X-axis is the future time , and the Y-axis is the average hourly loss rate . The red dashed line is the cumulative running loss rate, representing the cost of delay. As time goes on, the scaling becomes thicker and thicker, and the additional energy consumption grows faster, so the curve is monotonically increasing. The purple dashed line is the amortized shutdown loss rate, representing the cost of impatience. If the shutdown is frequent, the fixed shutdown loss (such as labor cost and cleaning cost) will be high. The longer the running time, the thinner the fixed shutdown loss is amortized, so the curve is monotonically decreasing. The blue solid line is the total average loss rate, which is the superposition of the red dashed line and the purple dashed line, showing a perfect U-shaped curve. Draw a black vertical line at the lowest point of the blue U-shaped curve, and the X-axis coordinate corresponding to the black vertical line is the optimal cleaning time point solved in step S4 . This intuitively shows that the present invention has found the best technical balance point between "energy consumption loss of running a little longer" and "labor loss of shutting down now".

[0090] In this way, by constructing the total average loss rate function and solving its minimum value, the present invention can abandon passive or fixed cycle maintenance strategies and achieve optimal maintenance decisions based on the dynamic balance of running energy consumption and shutdown loss.

[0091] The application provides a specific embodiment of an evaporation crystallization process optimization system based on multi-source data fusion:

[0092] The evaporation crystallization process optimization system based on multi-source data fusion comprises the following modules:

[0093] A data acquisition module is configured to acquire multi-source time series data of the evaporation crystallization system and construct an evaporation load index for representing a comprehensive load of a real-time working condition; the multi-source time series data at least comprises feed flow, feed concentration, steam compressor related parameters and a real-time observed heat exchange coefficient;

[0094] A model training and prediction module is configured to train a working condition decoupling model based on historical health data of the evaporation crystallization system, and predict a theoretical health heat exchange coefficient that the system should have when in a healthy state under a current working condition according to real-time acquired working condition variables by using the working condition decoupling model;

[0095] A scaling extraction and prediction module is configured to extract a pure scaling loss signal by calculating a residual error between the theoretical health heat exchange coefficient and the real-time observed heat exchange coefficient, and predict a scaling loss trend in a future time period based on historical data of the pure scaling loss signal;

[0096] An optimization decision module is configured to construct a total average loss rate function representing a balance relationship between energy loss and shutdown opportunity loss of the system, and solve a minimum value of the total average loss rate function according to the scaling loss trend to determine an optimal cleaning time point of the evaporation crystallization system, so as to realize dynamic optimization of the evaporation crystallization process.

[0097] It should be noted that the processing procedures of the data acquisition module, the model training and prediction module, the scaling extraction and prediction module and the optimization decision module respectively correspond to the processing procedures of steps S1, S2, S3 and S4 in each embodiment of the above evaporation crystallization process optimization method based on multi-source data fusion, and will not be described in detail here.

[0098] Although the present application has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made thereto without departing from the spirit and scope of the present application.

Claims

1. An evaporation crystallization process optimization method based on multi-source data fusion, characterized in that, The method comprises the following steps: Collecting multi-source time series data of the evaporation crystallization system, and constructing an evaporation load index for representing a comprehensive load of a real-time working condition; the multi-source time series data at least includes feed flow, feed concentration, steam compressor related parameters and real-time observed heat exchange coefficient; The calculation formula of the evaporation load index is: ; where, is the evaporative load index at time, and are the feed flow rate and feed concentration at time, and are the historical steady-state average flow rate and historical steady-state average concentration, and are the sensitivity weight factors for flow rate and concentration, respectively, is the natural exponential function; Based on historical health data of the evaporative crystallization system, a working condition decoupling model is trained, including: a gated recurrent unit (GRU) network model is constructed, containing variables respectively: evaporative load index , steam compressor outlet pressure , and compressor operating current , real-time working condition variables and data of the previous N-1 time points are input into the model, and the model outputs a predicted value ; using the working condition decoupling model, the theoretical health heat exchange coefficient that the system should have in a healthy state under the current working condition is predicted according to the real-time collected working condition variables; extracting a pure fouling loss signal by calculating a residual between the theoretical health heat exchange coefficient and the real-time observed heat exchange coefficient and predicting a fouling loss trend in a future time period based on historical data of the pure fouling loss signal ​ is a smooth, monotonically increasing curve that only reflects the true physical degree of fouling accumulation; A total average loss rate function representing a balance relationship between energy consumption loss and shutdown opportunity loss of the system is constructed, and a minimum value of the total average loss rate function is solved according to the scaling loss trend, so as to determine an optimal cleaning time point of the evaporation crystallization system, and realize dynamic optimization of the evaporation crystallization process.

2. The method for evaporation crystallization process optimization based on multi-source data fusion according to claim 1, characterized in that, The working condition decoupling model is a gated recurrent unit network model, and the historical health data is the multi-source time series data in a preset time period after each chemical cleaning of the evaporation crystallization system is completed.

3. The method for evaporation crystallization process optimization based on multi-source data fusion according to claim 1, characterized in that, The calculation formula of the pure scaling loss signal is: ; wherein, is the pure fouling loss signal at the moment, is the theoretical health heat transfer coefficient output by the working condition decoupling model, is the real-time observed heat transfer coefficient collected in real time.

4. The method for evaporation crystallization process optimization based on multi-source data fusion according to claim 3, characterized in that, The scaling loss trend in a future time period includes: fitting the pure fouling loss signal with a polynomial regression model historical data, and extrapolating to predict the fouling loss at future time instants .

5. The method for evaporation crystallization process optimization based on multi-source data fusion according to claim 4, characterized in that, The total average loss rate function is: ; wherein is a future cleaning time point to be determined, is the current time, is the accumulated running loss from to the time due to fouling accumulation, is the shutdown opportunity loss, is the fixed cleaning time, is the running time, is the average power in the healthy state.

6. The method for evaporation crystallization process optimization based on multi-source data fusion according to claim 5, characterized in that, The additional power consumption in the cumulative operational loss is which is calculated by historical data fitting or thermodynamic formula, represents the scaling loss as When the steam compressor is compared to the average power under the healthy state The additional power consumption required.

7. The multi-source data fusion based evaporative crystallization process optimization method of claim 5, wherein, The optimal cleaning time point is the time point that minimizes the value of the total average loss rate function value.

8. The multi-source data fusion based evaporative crystallization process optimization method of claim 1, wherein, The steam compressor related parameters at least include steam compressor outlet pressure and main steam compressor operating current.

9. A system for implementing the method for optimization of the evaporation crystallization process based on multi-source data fusion according to any one of claims 1-8, characterized in that, The method comprises the following modules: A data collection module is configured to collect multi-source time series data of the evaporation crystallization system, and construct an evaporation load index for representing a comprehensive load of a real-time working condition; the multi-source time series data at least includes feed flow, feed concentration, steam compressor related parameters and real-time observed heat exchange coefficient; A model training and prediction module is configured to train a working condition decoupling model based on historical health data of the evaporation crystallization system, and predict a theoretical health heat exchange coefficient that the system should have when in a healthy state under a current working condition by using the working condition decoupling model according to real-time collected working condition variables; A scaling extraction and prediction module is configured to extract a pure scaling loss signal by calculating a residual error between the theoretical health heat exchange coefficient and the real-time observed heat exchange coefficient, and predict a scaling loss trend in a future time period based on historical data of the pure scaling loss signal; An optimization decision module is configured to construct a total average loss rate function representing a balance relationship between energy consumption loss and shutdown opportunity loss of the system, and solve a minimum value of the total average loss rate function according to the scaling loss trend, so as to determine an optimal cleaning time point of the evaporation crystallization system, and realize dynamic optimization of the evaporation crystallization process.

Images