Intelligent material matching method and system for reducing furnace to improve polysilicon production efficiency
By building a material ratio prediction model and using machine learning and deep learning models to guide the material ratio of the reduction furnace, the problems of difficult quality control and low efficiency caused by manual experience in polysilicon production have been solved, and intelligent and efficient production has been achieved.
Patent Information
- Application Number
- CN202310565054.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-18
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2043-05-18
AI Technical Summary
In the existing polysilicon production, material addition relies on manual experience, which leads to difficult to control product quality, high energy consumption, low production efficiency, and difficulty in meeting actual needs.
A material ratio prediction model was built, and the machine learning model XGboost and the deep learning model LSTM were used to generate production condition prediction information based on the time series operation data of the reduction furnace system, guide the material ratio of the reduction furnace, and automatically add materials through the DCS system.
The intelligent level and production efficiency of polysilicon production have been improved, ensuring stable product quality and reducing energy consumption.
Smart Images
Figure CN116639696B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a material proportioning method and system, and in particular to an intelligent material proportioning method and system for a reduction furnace for improving the production efficiency of polysilicon. Background Art
[0002] A reduction furnace is a necessary equipment for polysilicon production. When using a reduction furnace for polysilicon production, under certain conditions, a certain molar ratio of high-purity trichlorosilane and high-purity hydrogen is introduced into the reduction furnace, and a physical and chemical reaction occurs on the surface of the inverted U-shaped silicon core rod to generate rod-shaped polysilicon.
[0003] In traditional polysilicon production, material addition is mostly based on manual experience and judgment, that is, material addition is achieved by adding a list of ingredients to the DCS (Distributed Control System). After the materials are added, the status of the reduction furnace is monitored in real time to achieve high-quality polysilicon production.
[0004] From the above description, it can be seen that when using a reduction furnace for polysilicon production, since materials are added mostly relying on manual experience and judgment, there will be problems with personnel allocation, uncontrollable product quality, high energy consumption, etc., which makes it difficult to meet the actual polysilicon production needs. Summary of the Invention
[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and system for intelligent proportioning of materials in a reduction furnace to improve the production efficiency of polysilicon. The method and system can effectively realize the material proportioning during polysilicon production, improve the intelligence level and production efficiency during polysilicon production, and be safe and reliable.
[0006] According to the technical solution provided by the present invention, a reduction furnace material intelligent proportioning method for improving polysilicon production efficiency comprises:
[0007] For a reduction furnace system used in polysilicon production, a material ratio prediction model for the reduction furnace system is constructed, wherein:
[0008] When using a reduction furnace system for polysilicon production, obtaining current production condition status information, including actual trichlorosilane flow rate, trichlorosilane valve opening, actual hydrogen flow rate, hydrogen valve opening, actual furnace cooling water flow rate, furnace cooling water valve opening, and actual values of six-loop currents used to heat silicon rods;
[0009] For the production condition status information obtained above, the material ratio prediction model is used to generate the production condition prediction information at the next moment, so as to use the production condition prediction information to guide the reduction furnace material ratio of the reduction furnace system during polysilicon production, wherein the production condition prediction information includes the trichlorosilane flow prediction value, the hydrogen flow prediction value and the six-ring current prediction value.
[0010] When building a material ratio prediction model, it includes:
[0011] Provides machine learning model XGboost and deep learning model LSTM;
[0012] Acquire time series operation data of the reduction furnace system when producing polysilicon, and generate a training data set based on the acquired time series operation data, wherein the training data set includes production condition status information at multiple different moments;
[0013] The training datasets generated above are used to train the machine learning model XGboost and the deep learning model LSTM respectively, so as to generate a material ratio prediction model after reaching the target training state;
[0014] During operation, the machine learning model XGboost is used to generate trichlorosilane flow prediction values and hydrogen flow prediction values within the production condition prediction information; at the same time, the deep learning model LSTM is used to generate the six-ring current prediction values within the production condition prediction information.
[0015] The acquired time series operation data is preprocessed, and after the preprocessing, the types of training data in the time series operation data are determined based on feature analysis, where:
[0016] The determined training data types include trichlorosilane flow rate, trichlorosilane valve opening, hydrogen flow rate, hydrogen valve opening, furnace cooling water flow rate, furnace cooling water valve opening, and six-loop current values used to heat silicon rods;
[0017] For the time series operation data, operation data information corresponding to the determined training data type is selected to form a training data set based on the selected operation data information, wherein,
[0018] Feature analysis methods include variance filtering.
[0019] When using variance filtering to perform feature analysis on time series data, the following steps are included:
[0020] Determine the variance of each feature data in the time series running data, and eliminate the corresponding feature data based on the variance of the feature data;
[0021] For the remaining feature data, the correlation coefficient of any two feature data is determined, and the feature data with a correlation coefficient greater than a correlation coefficient threshold is used as the training data type.
[0022] Preprocessing of time series data, including missing value processing, outlier processing and / or standardization processing, where:
[0023] When dealing with missing values, the missing value processing methods include missing deletion method and / or missing filling method
[0024] When processing outliers, the outlier processing methods include quartile method, absolute median method, outlier deletion method or replacement method.
[0025] For the deletion deletion method, the threshold value of the deletion deletion method is configured to be 0.8; for the deletion filling method, the threshold value of the deletion filling method is configured to be 0.05.
[0026] When training the machine learning model XGboost, the symmetric mean absolute percentage error SMAPE is used to measure the training status of the machine learning model XGboost, where
[0027] When the symmetric mean absolute percentage error SMAPE is not higher than 5%, the machine learning model XGboost reaches the target training state.
[0028] For the symmetric mean absolute percentage error SMAPE, we have:
[0029]
[0030] Among them, n is the total amount of data in the time series running data, t is the time in the time series running data, F t A is the data type of the training data at time t for the time series running data. t Outputs the predicted value at time t for the machine learning model XGboost.
[0031] When the target training state is reached, a basic model for material ratio prediction is generated first;
[0032] For the production condition status information at a current moment, the material ratio prediction basic model is used to generate basic production condition prediction information, the generated basic production condition prediction information is used to guide the reduction furnace material ratio, and the polysilicon production status under the guided reduction furnace material ratio state is verified, wherein,
[0033] When verifying the polysilicon production status, if the polysilicon production verification indicators pass and no abnormal conditions occur, the material ratio prediction basic model is configured as the material ratio prediction model. Otherwise, the training status of the machine learning model XGboost and the deep learning model LSTM is adjusted until the material ratio prediction model is obtained.
[0034] No abnormal conditions occur, including no atomization condition.
[0035] A reduction furnace material intelligent proportioning system for improving polysilicon production efficiency includes a reduction furnace system and a material proportioning processor. The reduction furnace system includes a reduction furnace device and a DSC system for controlling polysilicon production in the reduction furnace device, wherein:
[0036] For the reduction furnace device, the material ratio processor uses the above method to generate production condition prediction information, and loads the production condition prediction information into the DCS system, so that the DCS system controls the reduction furnace device to produce polysilicon under the production condition prediction information.
[0037] The advantages of the present invention are as follows: the material ratio prediction model is used to generate the production condition prediction information at the next moment for the above-obtained production condition status information, so as to use the production condition prediction information to guide the reduction furnace material ratio of the reduction furnace system during polysilicon production, wherein the production condition prediction information includes the trichlorosilane flow prediction value, the hydrogen flow prediction value and the six-ring current prediction value, that is, it can effectively realize the material ratio during polysilicon production, and improve the intelligence level and production efficiency during polysilicon production. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 This is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The present invention will be further described below with reference to specific drawings and embodiments.
[0040] In order to effectively realize the material ratio during polysilicon production and improve the intelligence level and production efficiency during polysilicon production, a reduction furnace material intelligent ratio method for improving polysilicon production efficiency is provided. In one embodiment of the present invention, the reduction furnace material intelligent ratio method includes:
[0041] For a reduction furnace system used in polysilicon production, a material ratio prediction model for the reduction furnace system is constructed, wherein:
[0042] When using a reduction furnace system for polysilicon production, obtaining current production condition status information, including actual trichlorosilane flow rate, trichlorosilane valve opening, actual hydrogen flow rate, hydrogen valve opening, actual furnace cooling water flow rate, furnace cooling water valve opening, and actual values of six-loop currents used to heat silicon rods;
[0043] For the production condition status information obtained above, the material ratio prediction model is used to generate the production condition prediction information at the next moment, so as to use the production condition prediction information to guide the reduction furnace material ratio of the reduction furnace system during polysilicon production, wherein the production condition prediction information includes the trichlorosilane flow prediction value, the hydrogen flow prediction value and the six-ring current prediction value.
[0044] Specifically, a reduction furnace system generally includes a reduction furnace device and a DCS system for controlling the polysilicon production process of the reduction furnace device. The reduction furnace device and DCS system can be used with commonly used polysilicon production equipment, and the specific method and process for implementing polysilicon production are consistent with existing methods. As can be seen from the above description, the existing technology manually adds ingredients to the DCS system to achieve material addition. This manual addition method results in low polysilicon production efficiency and a low degree of automation, making it difficult to meet actual production needs.
[0045] In one embodiment of the present invention, for a certain reduction furnace system, it is necessary to construct a material ratio prediction model based on the reduction furnace system, and the addition of material ratio can be achieved through the constructed material ratio prediction model.
[0046] When working, it is generally necessary to determine the production condition status information at the current moment, that is, the obtained production condition status information is used as the input information of the material ratio prediction model. The material ratio prediction model can generate / obtain the production condition prediction information at the next moment based on the production condition status information, that is, the production condition prediction information is the output of the material ratio prediction model.
[0047] In one embodiment of the present invention, the production condition status information includes at least the actual flow rate of trichlorosilane, the opening of the trichlorosilane valve, the actual flow rate of hydrogen, the opening of the hydrogen valve, the actual flow rate of furnace cooling water, the opening of the furnace cooling water valve, and the actual value of the six-loop current used to heat the silicon rods. Among them, the actual flow rate of trichlorosilane, the opening of the trichlorosilane valve, the actual flow rate of hydrogen, the opening of the hydrogen valve, the actual flow rate of furnace cooling water, the opening of the furnace cooling water valve, and the actual value of the six-loop current can be monitored by existing technical means.
[0048] The production condition prediction information includes the trichlorosilane flow rate prediction value, the hydrogen flow rate prediction value, and the six-loop current prediction value. Therefore, when the production condition prediction information is used to guide the reduction furnace system to adjust the reduction furnace material ratio during polysilicon production, the trichlorosilane flow rate prediction value is used as the reference value of the trichlorosilane flow rate in the DCS system, the hydrogen flow rate prediction value is used as the reference value of the hydrogen flow rate in the DCS system, and the six-loop current prediction value is used as the reference value of the silicon rod heating current. At this time, automatic addition of material ratios is achieved in the DCS, thereby improving the production efficiency of polysilicon.
[0049] For the reduction furnace system used in polysilicon production, the current moment and the next moment are related to the process and other conditions of polysilicon production using the reduction furnace system. The next moment specifically refers to the next process moment adjacent to the current moment. The specific conditions of the current moment and the next moment can be determined according to specific processing requirements.
[0050] In one embodiment of the present invention, when constructing a material ratio prediction model, the process includes:
[0051] Provides machine learning model XGboost and deep learning model LSTM;
[0052] Acquire time series operation data of the reduction furnace system when producing polysilicon, and generate a training data set based on the acquired time series operation data, wherein the training data set includes production condition status information at multiple different moments;
[0053] The training data sets generated above are used to train the machine learning model XGboost and the deep learning model LSTM respectively to generate a material ratio prediction model after reaching the target training state.
[0054] Specifically, the XGboost machine learning model is a commonly used machine learning model, and the LSTM deep learning model is a commonly used deep learning model. To generate the required production condition prediction information, the XGboost machine learning model and the LSTM deep learning model need to be trained.
[0055] The XGBoost machine learning model uses a GBDT (Gradient Boosting Decision Tree) decision tree. The number and depth of the trees affect the overall model structure. For implementation, existing commonly used models can be directly used for training. The LSTM (Long Short-Term Memory) deep learning model is a commonly used neural network-based model. For implementation, the activation function for multiple hidden layers within the LSTM deep learning model is set to leaky-relu (a nonlinear function, typically Tanh, Sigmoid, or Relu). Unlike the Relu function, the first half is set to αx instead of 0, with α set to 0.01. This increases the model's expressiveness and enables the network to better fit the data.
[0056] During training, the training data must be based on the current reduction furnace system's time-series operating data. This specifically refers to operating data during production hours. Using this data makes the trained material ratio prediction model more adaptable to the current reduction furnace system. In practice, the training dataset includes production condition status information at multiple different times. For details on this production condition status information, refer to the description above.
[0057] Reaching the target training state specifically refers to reaching the target training state after training the machine learning model XGboost and the deep learning model LSTM. The following details the target training states corresponding to the machine learning model XGboost and the deep learning model LSTM.
[0058] After reaching the target training state, the material ratio prediction model is formed using the machine learning model XGboost and the deep learning model LSTM. During operation, the machine learning model XGboost generates trichlorosilane flow rate prediction values and hydrogen flow rate prediction values within the production condition prediction information; simultaneously, the deep learning model LSTM generates the six-ring current prediction value within the production condition prediction information.
[0059] From the above description, when training the machine learning model XGboost and the deep learning model LSTM, the same training data set is used, and the technical means commonly used in the field of this technology are adopted to configure the corresponding output states of the machine learning model XGboost and the deep learning model LSTM. Therefore, after reaching the target training state, the machine learning model XGboost can generate the trichlorosilane flow prediction value and the hydrogen flow prediction value in the production condition prediction information, and the deep learning model LSTM can generate the six-ring current prediction value in the production condition prediction information.
[0060] In one embodiment of the present invention, the acquired time series operation data is preprocessed, and after the preprocessing, the training data type in the time series operation data is determined based on feature analysis and the balance principle in polysilicon production, wherein:
[0061] The determined training data types include trichlorosilane flow rate, trichlorosilane valve opening, hydrogen flow rate, hydrogen valve opening, furnace cooling water flow rate, furnace cooling water valve opening, and six-loop current values used to heat silicon rods;
[0062] For the time series operation data, operation data information corresponding to the determined training data type is selected to form a training data set based on the selected operation data information, wherein,
[0063] Feature analysis methods, including variance filtering;
[0064] The balance principles in polysilicon production include raw material balance, temperature balance and magnetic field balance.
[0065] Those skilled in the art will recognize that the specific process for polysilicon production involves introducing high-purity SiHCl3 (TCS) and H2 in a specific ratio into a chemical vapor deposition (CVD) reactor, commonly known as a "reduction furnace." A physical and chemical reaction occurs at 1080°C to 1200°C and 0.6 MPa, producing high-purity rod-shaped polysilicon and byproducts such as HCl, SiCl4 (STC), and SiH2Cl2 (DCS). The actual reaction is much more complex, with a single-pass SiHCl3 conversion rate of approximately 10%, requiring multiple cycles to gradually complete the reaction. Furthermore, the primary reaction is reversible at different temperatures, and all components of the reaction exhaust are recycled.
[0066] As can be seen from the polysilicon production process described above, the acquired time series operation data includes a large amount of feature data. Feature data specifically refers to a type of data related to polysilicon production. Therefore, after acquiring the time series operation data, it is necessary to determine the type of training data through technical means. In one embodiment of the present invention, the type of training data is determined based on feature analysis and the balance principle in polysilicon production. The determined type of training data is consistent with the type of input information of the material ratio prediction model.
[0067] After determining the type of training data, select all data information corresponding to the type of training data in the time series operation data, such as trichlorosilane flow rate, trichlorosilane valve opening, hydrogen flow rate, hydrogen valve opening, furnace cooling water flow rate, furnace cooling water valve opening and data values corresponding to the six-ring current value. At this time, the training data set can be formed.
[0068] In one embodiment of the present invention, when using the variance filtering method to perform feature analysis on time series operation data, the method includes:
[0069] Determine the variance of each feature data in the time series running data, and eliminate the corresponding feature data based on the variance of the feature data;
[0070] For the remaining feature data, the correlation coefficient of any two feature data is determined, and the feature data with a correlation coefficient greater than a correlation coefficient threshold is used as the training data type.
[0071] In specific implementations, after acquiring time series operational data, the number of feature data within the time series operational data can generally be determined. At this point, the variance of each feature data can be calculated using a commonly used variance formula. The variances of all feature data are sorted, and feature data with smaller variances are deleted. The variances of the feature data can be normalized to determine the normalized variances corresponding to all feature data. The data are then sorted in ascending order, and the top 5% to 10% of variances can be determined as having smaller variances. Alternatively, experience can be used to determine the feature data with smaller variances. Of course, other techniques can also be used to determine when the variance is smaller, and the specific method can be selected based on needs.
[0072] After removing the feature data with smaller variance, calculate the correlation coefficient between any two features of the remaining feature data. Then, compare all correlation coefficients with the correlation coefficient threshold. Only feature data with a correlation coefficient greater than the threshold is considered as training data. Generally, the correlation coefficient threshold is 0.5 to 0.7. The correlation coefficient between any two features can be calculated using a commonly used formula in the art, which will not be further illustrated here.
[0073] An analysis of the polysilicon production process shows that polysilicon production can be summarized as three major balances. Raw material balance refers to the balanced and stable ratio of SiHCl3 (TCS) and hydrogen; temperature balance refers to the balanced and stable chemical reaction conditions during polysilicon production; and magnetic field balance refers to the balanced and stable adsorption of silicon rods to the polysilicon generated after the reduction reaction. The three major balances in polysilicon production can be used to assist in the screening of feature data. Specifically, the screened data must meet the three major balances in polysilicon production.
[0074] In one embodiment of the present invention, the preprocessing of time series operation data includes missing value processing, outlier processing and / or standardization processing, wherein:
[0075] When dealing with missing values, the missing value processing methods include missing deletion method and / or missing filling method
[0076] When processing outliers, the outlier processing methods include quartile method, absolute median method, outlier deletion method or replacement method.
[0077] Typically, time series operational data is acquired from a DCS system. After acquisition, the data typically undergoes preprocessing to improve its reliability as a training dataset. Preprocessing of time series operational data typically includes missing value processing, outlier processing, and / or normalization. The specific preprocessing method depends on the state of the time series operational data.
[0078] When handling missing values, you can use the missing value deletion method and / or the missing value filling method. In specific implementation, read the time series data into a Dataframe using Python, and use the isna method in the Python pandas library to determine missing values. If the result is True, it indicates missing data, and the deletion method and filling method are used to handle it. Use the dropna method in the pandas library to delete, or use the fillna method in the pandas library to fill. For the missing value deletion method, set the threshold of the missing value deletion method to 0.8; for the missing value filling method, set the threshold of the missing value filling method to 0.05. Of course, other technical means can also be used to implement missing value determination. The specific method and process for handling missing values can be selected according to needs.
[0079] Interquartile range (IQR): The interquartile range (IQR), also known as the interquartile difference, is a method used in descriptive statistics to determine the difference between the third quartile and the first quartile. Like the variance and standard deviation, it indicates the dispersion of variables in statistical data and is a method for detecting outliers.
[0080] When using the quartile method to process outliers, specifically: for any feature data, remove the data that is greater than the maximum value or less than the minimum value in the time series running data, where:
[0081] Maximum value = Q3 + k(Q3 – Q1)
[0082] Minimum value = Q1-k(Q3–Q1)
[0083] Where Q1 is the first quartile, Q3 is the third quartile; k = 1.5 is a moderate outlier, and k = 3 is a severe outlier.
[0084] Median Absolute Deviation (MAD): Also known as Median Absolute Deviation (MAD), it is a method for detecting outliers by calculating the sum of the distances between each observation and the mean. When using the MAD method to handle outliers, specifically: for any feature data, calculate the distance between the feature data and the mean to detect outliers and remove them. The steps are as follows:
[0085] Step 1: Calculate the median (X) of the feature data.
[0086] Step 2: Calculate the absolute deviation between the feature data and the median abs(X-median(X));
[0087] Step 3: Calculate the median of the absolute deviations obtained in step 2. Specifically, MAD = median(abs(X - median(X)));
[0088] Step 4: Divide the absolute deviation value obtained in step 2 by the median obtained in step 3 to obtain a set of distance values from the center of all feature data based on MAD: abs(X-median(X)) / MAD.
[0089] Regarding the outlier deletion or replacement method, the commonly used technical means in this technical field can be used for processing. The specific method of the outlier deletion or replacement method will not be given any examples here.
[0090] Normalization, specifically the scaling of data, involves scaling the data so that it falls within a specific, small range. This removes unit constraints from the data and converts it into dimensionless values, making it easier to compare and weight indicators of different units or magnitudes.
[0091] In one embodiment of the present invention, when training the machine learning model XGboost, the symmetric mean absolute percentage error SMAPE is used to measure the training status of the machine learning model XGboost, wherein:
[0092] When the symmetric mean absolute percentage error SMAPE is not higher than 5%, the machine learning model XGboost reaches the target training state.
[0093] Specifically, SMAPE is the symmetric mean absolute percentage error (SMAPE), which is a regression model evaluation indicator that aims to reflect the difference between the model's predicted value and the actual value.
[0094] In specific implementation, for the symmetric mean absolute percentage error SMAPE, we have:
[0095]
[0096] Among them, n is the total amount of data in the time series running data, t is the time in the time series running data, F t A is the data type of the training data at time t for the time series running data. tOutputs the predicted value at time t for the machine learning model XGboost.
[0097] For time series operation data, it can generally include n data that meet the time series characteristics, that is, the total amount of data is n, and each data includes a production condition status information and a production condition output information corresponding to the production condition status information. The production condition output information is the actual value of trichlorosilane flow, the actual value of hydrogen flow and the actual value of six-ring current.
[0098] From the above description, we can see that after training the machine learning model XGboost, we can use the machine learning model XGboost to predict the output, that is, we can get the predicted value A at time t t , using the predicted value A at time t t The data F of the training data type at time t t , you can calculate the symmetric mean absolute percentage error (SMAPE). When the SMAPE is no higher than 5%, the XGboost machine learning model training is considered to have reached the target training state. During training, if the SMAPE is greater than 5%, it is generally necessary to adjust the learning rate of the XGboost machine learning model until the SMAPE is no higher than 5%.
[0099] For the LSTM deep learning model, logarithmic loss is used to evaluate the training status. Logarithmic loss, also known as logistic loss or cross-entropy loss, is defined based on probability estimation. It is commonly used in (multi-nominal) logistic regression and neural networks, as well as some variants of the expectation maximization algorithm, and can be used to evaluate the probabilistic output of a classifier.
[0100] There is usually no standard for using logarithmic loss to evaluate training status. Due to the characteristics of neural network models, the value of Logloss is generally made as small as possible. This can be achieved by adjusting parameters such as the number of network layers of the deep learning model LSTM.
[0101] In one embodiment of the present invention, when the target training state is reached, a basic model for predicting material ratio is first generated;
[0102] For the production condition status information at a current moment, the material ratio prediction basic model is used to generate basic production condition prediction information, the generated basic production condition prediction information is used to guide the reduction furnace material ratio, and the polysilicon production status under the guided reduction furnace material ratio state is verified, wherein,
[0103] When verifying the polysilicon production status, if the polysilicon production verification indicators pass and no abnormal conditions occur, the material ratio prediction basic model is configured as the material ratio prediction model. Otherwise, the training status of the machine learning model XGboost and the deep learning model LSTM is adjusted until the material ratio prediction model is obtained.
[0104] No abnormal conditions occur, including no atomization condition.
[0105] After training with the training dataset, the trained model is typically validated. Generally, after reaching the target training state, the generated model can be a basic material ratio prediction model. Once this basic material ratio prediction model is obtained, for any current production condition status information, the basic material ratio prediction model can be used to generate basic production condition prediction information based on the current production condition status information.
[0106] When the basic production condition prediction information is used to guide the material ratio of the reduction furnace, technical means commonly used in this technical field are used to obtain polysilicon production verification indicators. Among them, the polysilicon production verification indicators generally include the difference between the predicted trichlorosilane flow rate and the actual trichlorosilane flow rate, the difference between the predicted hydrogen flow rate and the actual hydrogen flow rate, and the difference between the predicted six-loop current and the actual six-loop current. In addition, it is necessary to use the tail gas temperature of the reduction furnace for auxiliary verification. The tail gas temperature is the temperature of the exhaust gas from the reduction furnace. The specific value of the tail gas temperature can be obtained by means of a temperature sensor, etc., wherein the tail gas temperature is the gas temperature when polysilicon is produced under the predicted trichlorosilane flow rate, the predicted hydrogen flow rate, and the predicted six-loop current.
[0107] During verification, if there are no significant anomalies between the predicted and actual trichlorosilane flow rates, the predicted and actual hydrogen flow rates, and the predicted and actual six-ring current values, and if the tail gas temperature is stable and shows no significant anomalies, the polysilicon production verification indicators are considered to have passed. "No significant anomalies" generally means no significant increase or decrease in the values, and the corresponding difference values and tail gas temperature remain stable.
[0108] When polysilicon production verification indicators pass verification, specifically when the polysilicon production verification indicators are excellent and no abnormal conditions occur, generally refers to the absence of atomization. During the polysilicon production process, atomization may occur in the following situations: an imbalance in the ratio of trichlorosilane feed to hydrogen, excessive manual adjustment of current and hydrogen pressure, feeding near the maximum feed value, excessive hydrogen, and excessive production of SiHCl3 by the reverse reaction. Atomization can affect the quality of polysilicon production.
[0109] If polysilicon production verification fails or an abnormality occurs, the training status of the XGboost and LSTM machine learning models needs to be adjusted. This means that training of these models needs to continue. As explained above, the XGboost machine learning model can be adjusted by adjusting the learning rate, for example. For the LSTM deep learning model, the number of network layers can be adjusted to achieve the desired training status.
[0110] In addition, when the polysilicon production verification index verification is passed and no abnormal state occurs, the current timing operation data can be used as the subsequent training data set. Otherwise, the current timing operation data should be eliminated to avoid affecting the training status.
[0111] In summary, a reduction furnace material intelligent proportioning system for improving polysilicon production efficiency can be obtained. In one embodiment of the present invention, the system includes a reduction furnace system and a material proportioning processor. The reduction furnace system includes a reduction furnace device and a DSC system for controlling the polysilicon production of the reduction furnace device, wherein:
[0112] For the reduction furnace device, the material ratio processor uses the above method to generate production condition prediction information, and loads the production condition prediction information into the DCS system, so that the DCS system controls the reduction furnace device to produce polysilicon under the production condition prediction information.
[0113] Specifically, the reduction furnace system can refer to the above description, and the material ratio processor can adopt the existing commonly used microprocessor form, which can be selected according to needs. The method and process of the material ratio processor generating production condition prediction information can refer to the above description and will not be repeated here.
Claims
1. A method for intelligently proportioning materials in a reduction furnace to improve polysilicon production efficiency, characterized in that: The reduction furnace material intelligent proportioning method includes: For a reduction furnace system used in polysilicon production, a material ratio prediction model for the reduction furnace system is constructed, wherein: When using a reduction furnace system for polysilicon production, obtaining current production condition status information, including actual trichlorosilane flow rate, trichlorosilane valve opening, actual hydrogen flow rate, hydrogen valve opening, actual furnace cooling water flow rate, furnace cooling water valve opening, and actual values of six-loop currents used to heat silicon rods; Based on the obtained production condition status information, a material ratio prediction model is used to generate production condition prediction information at the next moment, so as to use the production condition prediction information to guide the reduction furnace material ratio of the reduction furnace system during polysilicon production, wherein the production condition prediction information includes a trichlorosilane flow rate prediction value, a hydrogen flow rate prediction value, and a six-loop current prediction value; When building a material ratio prediction model, it includes: Provides machine learning model XGboost and deep learning model LSTM; Acquire time series operation data of the reduction furnace system when producing polysilicon, and generate a training data set based on the acquired time series operation data, wherein the training data set includes production condition status information at multiple different moments; The training datasets generated above are used to train the machine learning model XGboost and the deep learning model LSTM respectively, so as to generate a material ratio prediction model after reaching the target training state; During operation, the machine learning model XGboost is used to generate trichlorosilane flow prediction values and hydrogen flow prediction values within the production condition prediction information. At the same time, the deep learning model LSTM is used to generate the six-ring current prediction value within the production condition prediction information. The acquired time series operation data is preprocessed, and after preprocessing, the training data types in the time series operation data are determined based on feature analysis and the balance principle in polysilicon production, where: The determined training data types include trichlorosilane flow rate, trichlorosilane valve opening, hydrogen flow rate, hydrogen valve opening, furnace cooling water flow rate, furnace cooling water valve opening, and six-loop current values used to heat silicon rods; For the time series operation data, operation data information corresponding to the determined training data type is selected to form a training data set based on the selected operation data information, wherein, Feature analysis methods, including variance filtering; The balance principle in polysilicon production includes raw material balance, temperature balance and magnetic field balance. Raw material balance refers to the balanced and stable ratio of SiHCl3 and hydrogen, temperature balance refers to the balanced and stable chemical reaction conditions in the polysilicon production process, and magnetic field balance refers to the balanced and stable adsorption force of silicon rods on the polysilicon generated after the reduction reaction.
2. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to claim 1 is characterized in that: When using variance filtering to perform feature analysis on time series data, the following steps are included: Determine the variance of each feature data in the time series running data, and eliminate the corresponding feature data based on the variance of the feature data; For the remaining feature data, the correlation coefficient of any two feature data is determined, and the feature data with a correlation coefficient greater than a correlation coefficient threshold is used as the training data type.
3. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to claim 1 is characterized in that: Preprocessing of time series data, including missing value processing, outlier processing and / or standardization processing, where: When dealing with missing values, the missing value processing methods include missing deletion method and / or missing filling method When processing outliers, the outlier processing methods include quartile method, absolute median method, outlier deletion method or replacement method.
4. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to claim 3 is characterized in that: For the deletion deletion method, the threshold value of the deletion deletion method is configured to be 0.8; for the deletion filling method, the threshold value of the deletion filling method is configured to be 0.
05.
5. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to any one of claims 1 to 4, characterized in that: When training the machine learning model XGboost, the symmetric mean absolute percentage error SMAPE is used to measure the training status of the machine learning model XGboost, where When the symmetric mean absolute percentage error SMAPE is not higher than 5%, the machine learning model XGboost reaches the target training state.
6. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to claim 5 is characterized in that: For the symmetric mean absolute percentage error SMAPE, we have: Among them, n is the total amount of data in the time series running data, t is the time in the time series running data, F t A is the data type of the training data at time t for the time series running data. t Outputs the predicted value at time t for the machine learning model XGboost.
7. The intelligent ratio matching method for reducing furnace materials for improving polysilicon production efficiency according to any one of claims 1 to 4, characterized in that: In the target training state, the basic model for material ratio prediction is generated first; For the production condition status information at a current moment, the material ratio prediction basic model is used to generate basic production condition prediction information, the generated basic production condition prediction information is used to guide the material ratio of the reduction furnace, and the polysilicon production status under the guided reduction furnace material ratio state is verified, wherein, When verifying the polysilicon production status, if the polysilicon production verification indicators pass and no abnormal conditions occur, the material ratio prediction basic model is configured as the material ratio prediction model. Otherwise, the training status of the machine learning model XGboost and the deep learning model LSTM is adjusted until the material ratio prediction model is obtained. No abnormal conditions occur, including no atomization condition.
8. An intelligent material proportioning system for a reduction furnace for improving polysilicon production efficiency, characterized by: It includes a reduction furnace system and a material ratio processor, wherein the reduction furnace system includes a reduction furnace device and a DSC system for controlling the polysilicon production of the reduction furnace device, wherein: For the reduction furnace device, the material ratio processor uses the method of any one of claims 1 to 7 to generate production condition prediction information, and loads the production condition prediction information into the DCS system, so that the DCS system controls the reduction furnace device to produce polysilicon under the production condition prediction information.