A Prediction Method for the Reaction Endpoint of a Polysilicon Reduction Furnace Based on Deep Learning

Abstract

The present invention relates to the technical field of polysilicon production, and discloses a method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning. The method includes the following steps: sampling the operating parameters of the reduction furnace at regular intervals, and collecting the proportion of cauliflower material after the end of each production cycle; after each sampling of the operating parameters of the reduction furnace, calculating the silicon rod resistance and the silicon rod energization power; calculating the total energy consumption and the total trichlorosilane consumption from the start of production to the end of the reaction, as well as the resistance change rate at the end of the reaction; constructing and training an MLP neural network model through the data of part of the production cycles with the lowest proportion of cauliflower material, and automatically controlling the reaction end time point according to the shortest reaction time, the longest reaction time, and the set value of the proportion of cauliflower material.

Description

Technical Field

[0001] The present invention relates to the technical field of polysilicon production, and specifically relates to a method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning. Background Art

[0002] Polysilicon, as one of the important raw materials for producing photovoltaic devices, is the cornerstone of the photovoltaic industry. In recent years, with the development of the new energy industry, the production of polysilicon has received increasing attention. The current mainstream method for producing polysilicon is the improved Siemens method. This method utilizes the principle of chemical vapor deposition. When the temperature of the reduction furnace is around 1100 degrees Celsius, hydrogen and trichlorosilane undergo a vapor deposition reaction on a pre-treated silicon rod to form polysilicon. When the reaction is approaching the end, due to the large amount of polysilicon deposited on the surface of the silicon rod, the local temperature becomes too high. Continuing the reaction at this time will result in poor quality of the subsequently produced silicon, and this poor-quality silicon is called cauliflower material. Therefore, in the production process of polysilicon, it is necessary to timely determine the end point of the reaction and stop the reaction to improve the utilization efficiency of raw materials and reduce costs.

[0003] Currently, the production of polysilicon reduction furnaces mainly determines the reaction end point by manual observation, and this method has the disadvantages of poor accuracy and low efficiency. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning.

[0005] To solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning, comprising the following steps:

[0007] Step 1: Sample the operating parameters of the reduction furnace every time interval T to obtain production cycle data; the production cycle data includes the voltage U and current I passing through the silicon rod, the running time of the reduction furnace, the flow rate of trichlorosilane, the temperature inside the reduction furnace, and the temperature of the reduction furnace tail gas.

[0008] Step 2: Collect the proportion of cauliflower material in the polysilicon produced after each production cycle ends.

[0009] Step 3: After each sampling of the operating parameters of the reduction furnace, calculate the silicon rod resistance R through the formula R = U / I, and calculate the silicon rod power-on power P through the formula P = UI.

[0010] Step 4: Calculate the total energy consumption from the start of production to the end of the reaction The total consumption of trichlorosilane And the resistance change rate at the end of the reaction Among them, P i is the energization power of the silicon rod calculated after the i-th sampling, V i is the trichlorosilane flow rate at the i-th sampling, N is the total number of samplings, R i is the silicon rod resistance calculated after the i-th sampling, and m is the time length to be considered when calculating the resistance change rate;

[0011] Step Five: Denote the production cycle data of the part with the lowest cauliflower material ratio as production cycle data A. Construct an mlp neural network model through the production cycle data A and train it. The input variables of the mlp neural network model include the total energy consumption E, the total trichlorosilane consumption Q, the silicon rod resistance at the end of the reaction, the running time of the reduction furnace, the resistance change rate k at the end of the reaction, the temperature inside the reduction furnace at the end of the reaction, and the tail gas temperature of the reduction furnace at the end of the reaction. The output variable is the cauliflower material ratio;

[0012] Step Six: Denote the shortest reaction time in the production cycle data A as t min and the longest reaction time is t max ; In the actual production process, when the reaction time is greater than t min , and the cauliflower material ratio predicted by the mlp neural network model is lower than the set ratio value, then control the reaction to continue; when the reaction time is greater than t max , or when the cauliflower material ratio predicted by the mlp neural network model is greater than or equal to the set ratio value, then control the reaction to stop.

[0013] Specifically, the set ratio value is determined according to the difference between the current cauliflower material price and the non-cauliflower material price: when the difference between the cauliflower material price and the non-cauliflower material price is less than or equal to , the set ratio value is 40%, otherwise the set ratio value is 30%; where M is the mass of the finished polysilicon, t1 is the time required to produce polysilicon with a mass of M at a 40% cauliflower material ratio, t2 is the time required to produce polysilicon with a mass of M at a 30% cauliflower material ratio, I1 is the average value of the current required for production at a 40% cauliflower material ratio, I2 is the average value of the current required for production at a 30% cauliflower material ratio, P is the price of production electricity, C is the price of raw materials required for polysilicon production, w1 is the raw material conversion rate for production at a 40% cauliflower material ratio, and w2 is the raw material conversion rate for production at a 30% cauliflower material ratio.

[0014] Specifically, in Step Five, the production cycle data A is the production cycle data of the first 30% with the lowest cauliflower material ratio.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are:

[0016] The present invention provides a method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning. This method automatically determines the optimal furnace shutdown time according to the running time and operating status of the reduction furnace, and takes the maximization of economic benefits as the goal for determining the shutdown time after considering production energy consumption and the proportion of dense and non-dense materials in the product. Compared with traditional methods, this method does not rely on manual experience and observation, has higher accuracy, faster response, and can reduce labor costs at the same time; at the same time, this method allows the shutdown time to vary flexibly within a certain range according to the actual situation and can adapt to some unexpected situations in the production process. In addition, the judgment basis of this method will target changes in the market situation. This method can automatically determine a better proportion of dense and non-dense materials when shutting down the furnace to maximize the economic benefits of production enterprises, while traditional methods will not be adjusted according to the market situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of the prediction method in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] As Figure 1 shown, the present invention provides a method for predicting the reaction end point of a polysilicon reduction furnace based on deep learning, including the following steps:

[0020] S1: Sampling the operating parameters of the reduction furnace every time interval T to obtain production cycle data; the production data of the reduction furnace includes the voltage U / and current I passing through the silicon rod, the running time of the reduction furnace, the flow rate of trichlorosilane, the temperature inside the reduction furnace, and the temperature of the exhaust gas of the reduction furnace; in this embodiment, the time interval T is taken as 1 minute. In some embodiments, the time interval T can also be other set values such as half a minute, 2 minutes, 3 minutes, etc., which are mainly set according to the production cycle, sampling data volume requirements, etc. Generally, each production cycle is 7-10 days, which is set in advance according to the reduction furnace attributes and production requirements.

[0021] S2: Collecting the proportion of cauliflower materials in the polysilicon produced after each production cycle ends. The determination of the proportion of cauliflower materials usually includes manual evaluation and / or computer automatic recognition evaluation, and the general range is 15-30%.

[0022] S3: After each sampling of the operating parameters of the reduction furnace, calculating the resistance R of the silicon rod for this time through the formula R = U / I, and calculating the power P of the silicon rod for this time through the formula P = UI.

[0023] S4: Calculating the total energy consumption from the start of production to the end of the reaction Total consumption of trichlorosilane and the resistance change rate at the end of the reaction Where P i is the power of the silicon rod after the i-th sampling, V i is the flow rate of trichlorosilane at the i-th sampling, N is the total number of samplings, R i is the resistance at the i-th sampling, and m is the time length to be considered when calculating the resistance change rate; in this embodiment, generally m is taken as 10.

[0024] S5: Through the production cycle data of the lowest 30% of the cauliflower material ratio (denoted as production cycle data A), set the label value of the production cycle data A to 1, and set the label values of other production cycle data to 0; construct and train an mlp neural network model through the production cycle data A. The input variables of the mlp neural network model are the total energy consumption E, the total consumption of trichlorosilane Q, the resistance of the silicon rod at the end of the reaction, the operation time of the reduction furnace, the resistance change rate k at the end of the reaction, the temperature inside the reduction furnace at the end of the reaction, and the tail gas temperature of the reduction furnace at the end of the reaction, and the output is the proportion of cauliflower material in the polysilicon.

[0025] S6: Denote the shortest reaction time in the production cycle data A as t min and the longest reaction time as t max ; in the actual production process, when the reaction time is greater than t min , and the proportion of cauliflower material predicted by the mlp neural network model is lower than the set proportion value, then control the reaction to continue; when the reaction time is greater than t max , or when the proportion of cauliflower material predicted by the mlp neural network model is greater than or equal to the set proportion value, then control the reaction to stop. Through the setting of the longest reaction time t max and the shortest reaction time t min , even if abnormal disturbances occur in production and affect the model accuracy, the reaction can be controlled to end at a reasonable time point.

[0026] Specifically, before the start of a cycle of production, set the proportion of cauliflower material according to the current market conditions of cauliflower material and non-cauliflower material. When the price difference between cauliflower material and non-cauliflower material in the market is not large, the total output should be increased. At this time, the reaction time should be made as long as possible to increase the total output. Even if the proportion of cauliflower material will increase, in this embodiment, the proportion of cauliflower material can be set to 30%-35% at this time. When the market price difference between cauliflower material and non-cauliflower material is large, the reaction should be ended as early as possible to reduce the proportion of cauliflower material. In this embodiment, the proportion of cauliflower material is set at 20% at this time.

[0027] It is obvious to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, in any regard, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Accordingly, all changes that fall within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention, and any reference signs in the claims should not be construed as limiting the claims involved.

[0028] In addition, it should be understood that although this specification is described in terms of embodiments, not every embodiment contains only one independent technical solution. This narrative manner of the specification is merely for clarity. Those skilled in the art should view the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A prediction method for the reaction end point of a polysilicon reduction furnace based on deep learning, comprising the following steps: Step 1: Sampling the operating parameters of the reduction furnace every time interval T to obtain production cycle data; the production cycle data includes the voltage U and current I passing through the silicon rod, the running time of the reduction furnace, the flow rate of trichlorosilane, the temperature inside the reduction furnace, and the temperature of the tail gas of the reduction furnace. Step 2: Collecting the proportion of cauliflower material in the polysilicon produced after each production cycle ends. Step 3: After each sampling of the operating parameters of the reduction furnace, calculating the silicon rod resistance R through the formula R = U / I, and calculating the power P of the silicon rod energization through the formula P = UI. Step 4: Calculate the total energy consumption from the start of production to the end of the reaction Total consumption of trichlorosilane and the resistance change rate at the end of the reaction where P i is the power of the silicon rod calculated after the i-th sampling, V i is the flow rate of trichlorosilane at the i-th sampling, N is the total number of samplings, R i is the resistance of the silicon rod calculated after the i-th sampling, and m is the time length to be considered when calculating the resistance change rate; Step 5: Denote a part of the production cycle data with the lowest proportion of cauliflower material as production cycle data A, and construct and train an mlp neural network model through the production cycle data A. The input variables of the mlp neural network model include the total energy consumption E, the total consumption of trichlorosilane Q, the silicon rod resistance at the end of the reaction, the running time of the reduction furnace, the resistance change rate k at the end of the reaction, the temperature inside the reduction furnace at the end of the reaction, and the temperature of the tail gas of the reduction furnace at the end of the reaction, and the output variable is the proportion of cauliflower material. Step 6: Denote the shortest reaction time in production cycle data A as t min , and the longest reaction time as t max ; actual During the production process, when the reaction time is greater than t min , and when the proportion of the cauliflower material predicted by the MLP neural network model is lower than the set proportion value, the reaction is controlled to continue; when the reaction time is greater than t max , or when the proportion of the cauliflower material predicted by the MLP neural network model is greater than or equal to the set proportion value, the reaction is controlled to stop.

2. The prediction method for the reaction end point of a polysilicon reduction furnace based on deep learning according to claim 1, wherein The proportion setting value is determined according to the difference between the current price of cauliflower material and the price of non-cauliflower material: when the difference between the price of cauliflower material and the price of non-cauliflower material is less than or equal to , the proportion setting value is 40%; otherwise, the proportion setting value is 30%. Where M is the mass of the finished polysilicon, t1 is the time required to produce polysilicon with a mass of M according to the proportion of 40% cauliflower material, t2 is the time required to produce polysilicon with a mass of M according to the proportion of 30% cauliflower material, I1 is the average value of the current required for production according to the proportion of 40% cauliflower material, I2 is the average value of the current required for production according to the proportion of 30% cauliflower material, P is the price of production electricity, C is the raw material price required for producing polysilicon, w1 is the raw material conversion rate for production according to the proportion of 40% cauliflower material, and w2 is the raw material conversion rate for production according to the proportion of 30% cauliflower material.

3. The prediction method for the reaction end point of a polysilicon reduction furnace based on deep learning according to claim 1, wherein: In Step 5, the production cycle data A is the production cycle data of the first 30% with the lowest proportion of cauliflower material.

Images