A method for predicting local anode effect of a large aluminum electrolysis cell

Abstract

The application provides a large-scale aluminum electrolysis cell local anode effect prediction method, and belongs to the technical field of aluminum electrolysis. The method comprises the following steps: obtaining an anode rod current sequence, and performing fluctuation monitoring on the obtained anode rod current sequence; if fluctuation of the anode rod current is monitored, tracking prediction is performed on the anode rod current in the region where the anode rod is located and where no fluctuation occurs by using an ARIMA model, and tracking prediction is performed on the anode rod current in the region where the anode rod is located and where fluctuation has occurred by using an LSTM-GRU hybrid neural network, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit; and the local anode effect grade in the region is determined based on the tracking prediction results of the ARIMA model and the LSTM-GRU hybrid neural network. By using the application, the prediction accuracy of fluctuating current and the prediction accuracy of local anode effect can be improved, and the calculation amount is reduced.

Description

Technical Field

[0001] This invention relates to the field of aluminum electrolysis technology, and in particular to a method for predicting the local anodic effect in large aluminum electrolysis cells. Background Technology

[0002] In modern aluminum electrolytic plants, most anode effects occur due to the gradual diffusion of a localized decrease in alumina concentration to a decrease in alumina concentration throughout the entire cell. Detection methods for localized anode effects are still in their infancy. Measuring alumina concentration in different regions is challenging, requiring cell shell removal, electrolyte collection, and offline laboratory analysis. Simultaneous extraction of alumina concentration from multiple points is impossible, making it difficult to directly predict anode effects based on distributed alumina concentration. Furthermore, the difficulty in collecting distributed alumina concentration data makes accurate prediction of alumina concentration changes in different regions using data-driven methods also challenging. Therefore, many studies have primarily approached the issue from two perspectives: mechanistic analysis and data-driven approaches.

[0003] From the perspective of the mechanism of local anode effect: Ding Peilin et al. proposed an analysis of alumina concentration distribution during the feeding process based on finite element method (FEM) mechanism analysis. However, due to the long calculation time of the FEM algorithm, obtaining training data would require a long calculation period, making it infeasible for predicting the anode effect. WONG et al., through simulation of the equivalent circuit of an aluminum electrolysis cell, obtained the influence of fluoride bubbles on the cell voltage, and thus predicted the occurrence of local anode effect.

[0004] Analysis from a data-driven perspective: Some researchers, through analysis of distributed anode current, categorized anode effects into conventional anode effects and low-voltage anode effects, using a cell voltage threshold of 8V as the boundary. The detection of low-voltage anode effects was achieved by analyzing the non-uniformity of the anode rod current distribution, demonstrating the feasibility of using distributed anode rod current for anode effect prediction. CUI et al., through power spectrum analysis of distributed anode rod current, divided the aluminum electrolysis cell into six regions and determined the occurrence of local anode effects by setting current fluctuation thresholds. Invention patent (application number: CN201810078867.7) et al. developed a low-voltage anode effect prediction method based on the periodic average current deviation ratio by averaging the distributed anode current over time windows, and proposed a multi-point feeding strategy based on the probability distribution of low-voltage anode effect occurrence. However, since the threshold setting is fixed, the prediction rate decreases over time. Furthermore, when the threshold is set too high, missed detections occur; when the threshold is set too low, false alarms occur. Summary of the Invention

[0005] The embodiment of the application provides a large-scale aluminum electrolysis cell local anode effect prediction method, which can improve the prediction accuracy of fluctuating current and the prediction accuracy of local anode effect while reducing the calculation amount. The technical scheme comprises:

[0006] An anode rod current sequence is acquired, and fluctuation monitoring is performed on the acquired anode rod current sequence;

[0007] If it is monitored that the anode rod current fluctuates, an ARIMA model is used to track and predict the anode rod current in the region where the anode rod is located and does not fluctuate, and an LSTM-GRU hybrid neural network is used to track and predict the anode rod current in the region where the anode rod is located and has fluctuated, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit.

[0008] Based on the tracking and prediction results of the ARIMA model and the LSTM-GRU hybrid neural network, the local anode effect level in the region is determined.

[0009] Further, the acquired anode rod current sequence is represented as:

[0010] I anode =(i(t1),i(t2),i(t3),i(t4),…,i(t n ))

[0011] Wherein, I anode is the acquired anode rod current sequence, i(t n ) represents the anode rod current data of the t n th time collection point, and n is the time of collecting total anode rod current data.

[0012] Further, the fluctuation monitoring on the acquired anode rod current sequence comprises:

[0013] The acquired anode rod current sequence is divided by a fixed-length sliding window to obtain (n-m) / l+1 sub-time window current sequences, wherein n is the time of collecting total anode rod current data, m is the length of the sliding window, and l is the step length of the sliding window moving backward each time.

[0014] The slope of the jth sub-time window current sequence Y j is averaged to obtain k mean , and the average value c of each sub-time window current sequence is calculated, and the sliding window length m, the slope average value k mean , and the average value c constitute an array matrix X j (m, k mean , c), wherein 1≤j≤(n-m) / l+1.

[0015] the distance radius threshold d of the slope confidence interval of the jth sub-time window current sequence Y j j j is greater than the fluctuation threshold β, if yes, it is determined that the current sequence of the jth sub-time window fluctuates.

[0016] Further, the distance radius threshold d of the slope confidence interval of the jth sub-time window current sequence Y j j is expressed as:

[0017]

[0018]

[0019]

[0020] wherein, is the upper limit of confidence; θ j is the lower limit of confidence; is the slope mean value of the jth sub-time window current sequence, σ j is the mean square deviation of the slope of the jth sub-time window current sequence; Z represents the standard normal distribution; and α represents the confidence level.

[0021] Further, the tracking prediction of the anode guide rod current in the area where the anode guide rod is located without fluctuation by using the ARIMA model comprises:

[0022] tracking prediction of the anode guide rod current in the area where the anode guide rod is located without fluctuation by using the ARIMA model, outputting the unfluctuated anode guide rod current prediction value I 预测 , if the error between I 预测 and the actual current I 真实 is greater than the set threshold α, the tracked anode guide rod fluctuates.

[0023] Further, before tracking prediction of the anode guide rod current in the area where the anode guide rod is located with fluctuation by using the LSTM-GRU hybrid neural network, the method comprises:

[0024] defining the strong fluctuation current value of the anode as greater than or equal to 1.5 times the mean value of the 20 current sampling values before fluctuation, obtaining:

[0025]

[0026] wherein, ts is the time point of monitoring strong current fluctuation, I ts ​​​is a strong current fluctuation value; I t is a current sampling value;

[0027] The current data before the time point when the strong fluctuation current value I ts of the anode is exceeded is input into the LSTM-GRU hybrid neural network for training learning, wherein the trained LSTM-GRU hybrid neural network is used to track and predict the fluctuation of the anode guide rod current, to obtain the strong fluctuation current value I ts and the strong fluctuation occurrence time ts.

[0028] Further, the local anode effect level includes three levels of I, II and III; wherein,

[0029] I indicates that the strong fluctuation current prediction value of the fluctuation anode guide rod is greater than the set threshold value p, and the error of the non-fluctuation anode guide rod current prediction value I 预测 in the region and the actual current I 真实 is less than the set threshold value a; II indicates that the strong fluctuation current prediction value of the fluctuation anode guide rod is greater than the set threshold value p, and the error of the non-fluctuation anode guide rod current prediction value I 预测 in the region and the actual current I 真实 is greater than the set threshold value a;

[0030] III indicates that the strong fluctuation current prediction value I 预测 of the fluctuation anode guide rod and the actual current I 真实 is greater than the set threshold value p, and the non-fluctuation anode guide rod current prediction value in the region is greater than the set threshold value a.

[0031] The technical scheme provided by the embodiment of the present application has at least the following beneficial effects:

[0032] In the embodiment of the present application, the anode guide rod current sequence is obtained, and the fluctuation of the obtained anode guide rod current sequence is monitored; if the fluctuation of the anode guide rod current is monitored, the ARIMA model is used to track and predict the non-fluctuation anode guide rod current in the region where the anode guide rod is located, and the LSTM-GRU hybrid neural network is used to track and predict the fluctuation anode guide rod current in the region where the anode guide rod is located, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit; based on the tracking and prediction results of the ARIMA model and the LSTM-GRU hybrid neural network, the local anode effect level in the region is determined. In this way, according to the uneven distribution characteristics of the region current, according to the different current fluctuation, the ARIMA model and the LSTM-GRU hybrid neural network are used to classify and track the prediction of the anode guide rod current in the region, which improves the prediction accuracy of the fluctuation current and the prediction accuracy of the local anode effect, and reduces the calculation amount. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A schematic flowchart illustrating the method for predicting local anode effects in large aluminum electrolytic cells provided in this embodiment of the invention;

[0035] Figure 2 A schematic diagram illustrating the principle of sliding window movement provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the anode guide rod partitioning of a 400KA-class aluminum electrolytic cell provided in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram illustrating the process of tracking and predicting the current of the anode conductor in the region where the anode conductor is located without fluctuation, using an ARIMA model, as provided in an embodiment of the present invention.

[0038] Figure 5 This is a schematic diagram of the anode conductor current abnormality detection process provided in an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of the local anode effect prediction mechanism provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, this embodiment of the invention provides a method for predicting local anodic effects in large aluminum electrolysis cells, including:

[0042] S101, acquire the anode rod current sequence, and monitor the fluctuation of the acquired anode rod current sequence;

[0043] In this embodiment, the obtained anode conductor current sequence is represented as follows:

[0044] I anode =(i(t1),i(t2),i(t3),i(t4),…,i(t n )) (1)

[0045] Among them, I anode To obtain the anode conductor current sequence, i(t)n ) represents the anode guide rod current data of the t n th time collection point, and n is the time of collecting the total anode guide rod current data.

[0046] In this embodiment, the acquired anode guide rod current sequence is divided into a plurality of sub-time window current sequences by a sliding window, and data analysis is performed on each sub-time window current sequence for monitoring the fluctuation of the current, which can specifically include the following steps:

[0047] A1, as shown in Figure 2 , a fixed-length sliding window is used to slide and divide the acquired anode guide rod current sequence, obtaining (n-m) / l+1 sub-time window current sequences, wherein n is the time (in seconds) of collecting the total anode guide rod current data, m is the length of the sliding window, m << n, l is the step length of the sliding window moving backward each time, and the sliding window needs to slide (n-m) / l times;

[0048] A2, the average of the slope of the jth sub-time window current sequence Y j is obtained, that is, k mean , and the average value c of each sub-time window current sequence is calculated, and the sliding window length m, the average slope k mean , and the average value c form an array X j (m, k mean , c).

[0049] In this embodiment, it is assumed that one of the sub-time window current sequences is Y j (1≤j≤(n-m) / l+1), and the slope k i between any two adjacent data points in the sub-time window current sequence is calculated according to formula (2):

[0050]

[0051] wherein Y j represents the jth sub-time window current sequence, and j represents the sub-time window sequence number;

[0052] In the sub-time window current sequence Y j , there are m-1 slopes of adjacent data, the average of the slope of the sub-time window current sequence Y j is obtained, that is, k mean , and the average value c of the data of the jth sub-time window current sequence is calculated, and the sliding window length m, the average slope k mean , and the average value c form an array matrix X j (m, k mean , c).

[0053] A3, in order to better monitor the fluctuation of the current sequence, the distance radius threshold d of the slope confidence interval is set j The classification monitoring is performed to determine whether the current sequence generates fluctuation, specifically: first, the distance radius threshold d of the slope confidence interval of the jth sub-time window current sequence Y j j :

[0054]

[0055]

[0056]

[0057] wherein, is the upper limit of confidence; θ j is the lower limit of confidence; is the slope mean value of the jth sub-time window current sequence, σ j is the mean square deviation of the slope of the jth sub-time window current sequence; Z represents the standard normal distribution; and a represents the confidence level, which is used to improve or reduce the sensitivity of the confidence interval.

[0058] Then, it is judged whether d j is greater than the fluctuation threshold β, if yes, it is determined that the current sequence of the jth sub-time window generates fluctuation.

[0059] In the embodiment, according to the historical current fluctuation, the average value of 50 groups of historical fluctuation current data values greater than or equal to 1.5 times the average current data is selected as the fluctuation threshold β, if d j > β, it is determined that the current sequence of the sub-time window generates fluctuation, and the sub-time window number j of fluctuation is output.

[0060] In the embodiment, in order to further improve the detection rate, the density of the array matrix X j (m, k mean , c) and the adjacent ten (selecting the five before and five after the current window) sub-time windows is compared, if the more similar density of the current sequence is, the smaller the probability of generating fluctuation is. The array density of the sliding window and the distance between different sliding windows are related, taking the sub-time window objects X1 and X2 as an example, the distance thereof is defined as d(X1, X2):

[0061]

[0062] The fluctuation monitoring algorithm process is:

[0063] (1) Selecting the adjacent ten sub-time windows. ​

[0064] (2) For a sliding window, the reachable density lrd is calculated according to equation (7). k (X).

[0065]

[0066] In the formula, |N k (X)| represents the number of sub-time windows within the k-th distance (denoted as k-distance(X), where distance(X) is calculated using equation (6)) of the sub-window object X, N. k (X) is the set of sub-time window objects within the k-th distance neighborhood of sub-window object X, and o is N. k The elements in (X). The k-th distance neighborhood is defined as the set of objects in the selected sample space whose distance to object X is less than or equal to k-distance(X). reachdist(X,o) = max{k-distance(o),||Xo||} is the maximum value between the k-distance of object o in the sub-time window and the direct distance between object o and object X in the sub-time window.

[0067] (3) As in equation (8), select the mean of the reachability density of the ten nearest sub-time windows of object X and the object X. i The ratio of locally achievable densities to LOF k (X i The closer the value is to 1, the closer the density is, and the more stable the current sequence remains. The larger the value is than 1, the greater the difference between it and the mean density achievable in the nearest neighbor sub-time window, and the stronger the volatility.

[0068]

[0069] In this embodiment, the parameter LOF is used. k (X i To determine the intensity of current fluctuations.

[0070] To reduce the computational load for predicting the anode effect, such as Figure 3 As shown, taking a 400KA aluminum electrolysis cell as an example, it is divided into six areas according to the location and number of distributed feeding ports. Assuming that the current fluctuation of the anode guide rod B3 in area 1 occurs under the sliding window monitoring method, the current tracking and prediction of the eight guide rods in area 1 where the guide rod B3 is located will begin.

[0071] S102, if the fluctuation of the anode rod current is monitored, the ARIMA (Autoregressive Integrated Moving Average Model) model is used to track and predict the anode rod current in the region where the fluctuation does not occur, and the LSTM-GRU hybrid neural network is used to track and predict the anode rod current in the region where the fluctuation has occurred, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit.

[0072] 5-15 minutes before the local anode effect occurs, the current of one or several rods fluctuates, but the local alumina concentration of the part of the rod in contact with the electrolyte decreases (according to expert experience, when the alumina concentration is lower than 1.5%, the electrolyte balance will be broken, and the anode effect is easy to occur) requires a certain time to spread and diffuse to other similar adjacent rods. Tracking the current changes of the remaining anode rods in the same region that have not fluctuated will provide an important reference for anode effect prediction. Through comparative analysis, the ARIMA model can be used to track and predict the anode rod current in the region where the fluctuation does not occur, which has small calculation amount, is easy to operate, and has strong real-time performance.

[0073] In this embodiment, the ARIMA (p, q) order of the ARIMA model is determined to track the non-fluctuating current sequence and prevent strong current fluctuation.

[0074]

[0075] wherein t is the time point of current collection, I t is the collected anode rod current data; p and q are the order of the ARMA model; θ q is the model coefficient; the order is the difference order, and the model prediction quality is ensured; ε t is the error value in the current sequence; the order p and q are determined by the autocorrelation coefficient and the partial correlation function of the current sequence.

[0076] In this embodiment, as Figure 4As shown, the workflow of tracking and predicting the anode guide rod current in the region where the anode guide rod does not fluctuate using the ARIMA model includes: first, pre-processing the collected non-fluctuating anode guide rod current data, eliminating data with anode guide rod current greater than 30 KA, and 0 data not collected due to poor clamp contact; then, performing ADF (Augmented Dickey-Fuller, time series stationarity) test on the non-fluctuating 10 min anode guide rod current sequence, if the significant test statistic is less than the stationary threshold of 5%, it is determined to be stationary; if the significant test statistic is greater than the stationary threshold of 5%, perform the stationarity test again after first-order difference operation. The current sequence that passes the stationarity test is fitted using the ARIMA model, and the predicted value of the anode guide rod current I 预测 for the next 3 min is output, if the error between I 预测 and the actual current I 真实 is greater than the set threshold α, the tracked guide rod has anode current fluctuation, and the prediction accuracy of the local anode effect is improved.

[0077] In this embodiment, the anode strong fluctuation current value I ts is defined as 1.5 times the average of the 20 current sampling values before fluctuation, as shown in formula (10):

[0078]

[0079] Where ts is the time point when the strong current fluctuation is monitored, I ts is the strong current fluctuation value; I t is the current sampling value.

[0080] In this embodiment, I ts is used to determine the time point when the current fluctuation occurs, so that the current data before this time point (for example, the current data of the previous 30 minutes) is selected to train the LSTM-GRU hybrid neural network.

[0081] In this embodiment, in order to better improve the prediction accuracy of the local anode effect, and because the fluctuating anode guide rod current has strong instability, the tracking effect using the ARIMA model and the calculation amount cannot meet the real-time monitoring requirements, therefore, for the strong fluctuation of the current, in this embodiment, the LSTM-GRU hybrid neural network is used to track and predict the anode guide rod current in the region where the anode guide rod has fluctuated, and the local anode effect prediction time is advanced to 5-10 min, providing a theoretical basis for designing an anode effect self-extinguishing strategy.

[0082] In this embodiment, as shown in Figure 5 , by monitoring the anode strong fluctuation current value I tsThe current data 30 minutes before the time point is input into the LSTM-GRU hybrid neural network for training and learning, wherein the trained LSTM-GRU hybrid neural network is used to track and predict the anode rod current of the occurred fluctuation, to obtain strong fluctuation current value I ts and strong fluctuation occurrence time ts, to achieve early prediction of fluctuation current;

[0083] In S103, based on the tracking and prediction results of the ARIMA model and the LSTM-GRU hybrid neural network, the local anode effect level in the region is determined.

[0084] In this embodiment, combined with the electrolytic tank alumina concentration distribution, considering various working conditions of the anode rod current in the region, a fuzzy level classification of local anode effect prediction is proposed.

[0085] In this embodiment, as shown in Figure 6 , the prediction result of the ARIMA model and the prediction result of the LSTM-GRU hybrid neural network jointly determine the local anode effect prediction result, improving the prediction credibility; and in order to reduce the false positive rate, the local anode effect is fuzzy processed and divided into three levels: I, II, and III; wherein,

[0086] I represents that the strong fluctuation current prediction value of the fluctuation anode rod is greater than the set threshold value p, and the error of the current prediction value I 预测 of the non-fluctuation anode rod in the region and the actual current I 真实 (referred to as: true value) is less than the set threshold value a;

[0087] II represents that the strong fluctuation current prediction value of the fluctuation anode rod is greater than the set threshold value p, and the error of the current prediction value I 预测 of one or more non-fluctuation anode rods in the region and the actual current I 真实 is greater than the set threshold value a;

[0088] III represents that the strong fluctuation current prediction value of the fluctuation anode rod is greater than the set threshold value p, and the error of the current prediction value I 预测 of three or more non-fluctuation anode rods in the region and the actual current I 真实 is greater than the set threshold value a.

[0089] The large-scale aluminum electrolysis cell local anode effect prediction method provided by the embodiment of the present application obtains an anode rod current sequence, and performs fluctuation monitoring on the obtained anode rod current sequence; if it is monitored that the anode rod current fluctuates, the ARIMA model is used to track and predict the anode rod current in the region where the anode rod is located and where no fluctuation occurs, and the LSTM-GRU hybrid neural network is used to track and predict the anode rod current in the region where the anode rod is located and where fluctuation has occurred, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit; and the tracking and prediction results of the ARIMA model and the LSTM-GRU hybrid neural network are used to determine the local anode effect level in the region. In this way, according to the uneven current distribution characteristics of the region, the ARIMA model and the LSTM-GRU hybrid neural network are used to classify and track and predict the anode rod current in the region according to the different current fluctuation, so that the prediction accuracy of the fluctuating current and the prediction accuracy of the local anode effect are improved, and the calculation amount is reduced.

[0090] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of predicting local anode effect in a large-scale aluminum reduction cell, characterized by, The method comprises the following steps: acquiring an anode guide rod current sequence, performing fluctuation monitoring on the acquired anode guide rod current sequence; if fluctuation of the anode guide rod current is monitored, tracking prediction is performed on the anode guide rod current in the region where the anode guide rod is located and where no fluctuation occurs by using an ARIMA model, and tracking prediction is performed on the anode guide rod current in the region where the anode guide rod is located and where fluctuation has occurred by using an LSTM-GRU hybrid neural network, wherein ARIMA represents an autoregressive moving average model, LSTM represents a long short-term memory neural unit, and GRU represents a gated recurrent unit; based on the tracking prediction results of the ARIMA model and the LSTM-GRU hybrid neural network, a local anode effect level in the region is determined; wherein the fluctuation monitoring on the acquired anode guide rod current sequence comprises: the acquired anode guide rod current sequence is divided into (n-m) / l+1 sub-time window current sequences by using a fixed-length sliding window, wherein n is the time of collecting the total anode guide rod current data, m is the length of the sliding window, and l is the step length of the sliding window moving forward each time; The slope of the jth sub-time window current sequence Y j is averaged to obtain k mean , and the average value c of each sub-time window current sequence is calculated, and the sliding window length m, the slope average value k mean , and the average value c form an array matrix X j (m, k mean , c), wherein 1≤j≤(n-m) / l+1; set the jth sub time window current sequence Y j The distance radius threshold d of the slope confidence interval of j , determine whether d j is greater than the fluctuation threshold β, if yes, determine that the current sequence of the jth sub time window fluctuates; wherein the jth sub-time window current sequence Y j the distance radius threshold d of the slope confidence interval of j is expressed as: wherein is an upper confidence limit; θ j is a lower confidence limit; is the mean of the slopes of the jth sub-time window current sequence, σ j is the mean square deviation of the slopes of the jth sub-time window current sequence; Z denotes the standard normal distribution; and a denotes the confidence level.

2. The method of claim 1, wherein the method further comprises: the acquired anode guide rod current sequence is represented as: I anode = (i(t1), i(t2), i(t3), i(t4),..., i(t n )) where I anode is the acquired anode lead current sequence, i(t n ) represents the anode lead current data at the t n th time collection point, and n is the total number of time points at which anode lead current data is collected.

3. The method of claim 1, wherein the method further comprises: the tracking prediction of the anode guide rod current in the region where the anode guide rod is located and where no fluctuation occurs by using the ARIMA model comprises: The ARIMA model is used to track and predict the current of the anode rod that does not fluctuate within the region where the anode rod is located, and the predicted value I of the current of the anode rod without fluctuation is output for a certain period of time in the future. 预测 , if I 预测 With actual current I 真实 If the error exceeds the set threshold α, the current in the tracked anode conductor will fluctuate.

4. The method of claim 1, wherein the method further comprises: before the tracking prediction of the anode guide rod current in the region where the anode guide rod is located and where fluctuation has occurred by using the LSTM-GRU hybrid neural network, the method comprises: the anode strong fluctuation current value is defined as being greater than or equal to 1.5 times the average value of the 20 current sampling values before fluctuation, and is obtained as: wherein ts is the time point at which a strong current fluctuation is monitored, I ts is a strong current fluctuation value; I t is a current sample value; The time point before the current data exceeding the strong fluctuation current value I of the anode is monitored in advance ts The current data is input into the LSTM-GRU hybrid neural network for training learning, wherein the trained LSTM-GRU hybrid neural network is used to track and predict the fluctuation of the anode guide rod current, and the strong fluctuation current value I ts And the strong fluctuation occurrence time ts.

5. The method of claim 1, wherein the method further comprises: the local anode effect level comprises three levels of I, II and III; wherein, I indicates that the strong fluctuation current prediction value of the fluctuating anode guide rod is greater than the set threshold p, and the current prediction value I of the non-fluctuating anode guide rod in the region is less than the set threshold p 预测 The error of the actual current I 真实 is less than the set threshold a; II indicates that the strong fluctuation current prediction value of the fluctuating anode guide rod is greater than the set threshold p, and the current prediction value I of one or more than three non-fluctuating anode guide rods in the region is greater than the set threshold p 预测 The error of the actual current I 真实 is greater than the set threshold a; III represents the fluctuation anode guide rod strong fluctuation current prediction value I 预测 with the actual current I 真实 The error is greater than the set threshold value p, and the current prediction value of more than three non-fluctuation anode guide rods in the area is greater than the set threshold value a.

Images