Nickel matte temperature prediction method based on time sequence convolutional network and dynamic error compensation

Abstract

The invention relates to the technical field of industrial index prediction, in particular to a nickel matte temperature prediction method based on a time sequence convolutional network and dynamic error compensation, and aims to solve the problem of poor precision and robustness prediction of outlet matte temperature in the existing pyrometallurgy process. According to the method, through a Pearson's correlation coefficient and F-score fusion feature screening strategy, key input features are extracted from multi-dimensional parameters, and it is ensured that a model pays attention to most relevant input variables; on the basis of expansion causal convolution operation of a time sequence convolution network, capturing nonlinear coupling between parameters and time sequence dynamic characteristics, and generating an initial prediction result; dynamic compensation is realized by introducing an initial prediction error into an Xgboost model based on first-order and second-order time difference features, key input features are extracted through feature screening, nonlinear coupling and time sequence dynamic features among parameters are captured by using a TCN, an initial prediction result is generated and dynamic compensation is performed, and accurate prediction of industrial scene indexes is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial index prediction, and particularly relates to a nickel matte temperature prediction method based on a time series convolution network and dynamic error compensation. BACKGROUND

[0002] In the smelting pyrometallurgical process, the casting process plays an important role, and the core purpose is to complete the melting of various materials into the furnace under high temperature conditions and finally cast into anode plates. In this process, the matte temperature is an important indicator of the degree of reaction in the furnace and the stability of the process. Accurate monitoring and prediction of the matte temperature is of great significance for optimizing the casting operation, improving product quality and reducing energy consumption. In recent years, with the rapid development of deep learning, researchers have further applied it to metallurgical process modeling and soft measurement. For example, a multi-working-condition soft measurement method based on time series constraint fuzzy C-means clustering effectively improves the prediction accuracy and stability by combining mechanism and data-driven models. A combination model based on VMD-TCN-GRU realizes high-precision advance prediction of the blast furnace permeability index, effectively improving the accuracy and robustness of furnace condition monitoring. In industrial process modeling, error compensation is an important means to improve model performance and has been widely used. However, such prediction methods still have the problem of poor prediction accuracy and robustness of the outlet matte temperature in the pyrometallurgical process. SUMMARY

[0003] The present application provides a nickel matte temperature prediction method based on a time series convolution network and dynamic error compensation to solve the problems in the above background.

[0004] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0005] The nickel matte temperature prediction method based on a time series convolution network and dynamic error compensation comprises:

[0006] 1. A feature selection method based on Pearson correlation coefficient and F-score, which uses a fusion feature selection strategy of Pearson correlation coefficient (linear relationship) and F-score (nonlinear relationship) to compress the input dimension under the premise of ensuring the integrity of key information, and provides a high-quality, low-redundancy feature subset for subsequent index prediction.

[0007] 2. An inflation causal convolution operation based on a time series convolution network, which captures the nonlinear coupling and time series dynamic characteristics between parameters to generate an initial prediction result. After the key indicators are preliminarily predicted by the backbone network, the predicted values and true values of the training set and test set are collected to construct a training sample set of model errors.

[0008] 3. By introducing the initial prediction error into the Xgboost model based on first-order and second-order time difference features, the instantaneous rate of change and acceleration information of time series data are extracted through the training sample set, and the extracted features are used as input to the error compensation model to learn the evolution law of the error and achieve dynamic compensation.

[0009] 4. Calculate the local standard deviation through a sliding window, dynamically adjust the compensation weight and window length, distinguish between high and low volatility periods, increase the compensation intensity during high volatility periods, and make conservative corrections during low volatility periods to balance response speed and stability.

[0010] 5. By fusing the prediction results from different models, the final compensated prediction result is obtained through prediction value fusion: y final (t)=y base (t)+w(t)×e xgb (t), thereby improving prediction accuracy and reducing prediction error, where w(t) depends on local fluctuations in {w high ,w low The dynamic switching between these states reflects the differentiated handling of different fluctuations in the dynamic compensation strategy.

[0011] Furthermore, during the feature selection process, key influencing features were determined using Pearson and F-Score methods respectively, and for each candidate feature x... i Calculate its Pearson correlation coefficient r with the target variable y. i Take the absolute value |r i |And sort them in descending order, retaining the top 50% of the feature sets S. corr ; Calculate the F-Score (which measures the linear explanatory strength of a feature for y) for all features, and select the top 50% of features S. F .

[0012] Furthermore, during the feature selection process, a fused temporal convolutional network is used to calculate the intersection S. cand =S corr ∩S F The features in the dataset are evaluated based on their combined scores on the Pearson correlation coefficient and the F-Score, and the top 25% of features are retained to form set S. core .

[0013] Furthermore, during the feature selection process, a feature correlation coefficient matrix is ​​constructed to eliminate highly collinear features, ultimately yielding a deredundant set S. CFR Then, based on metallurgical mechanisms and production experience, the key variable S, such as "concentrate quantity," was analyzed. key All features are retained regardless of the selection results. The final feature set is defined as S. final =S CFR ∩S key .

[0014] Further, in the process of time series modeling, the time convolution network (TCN) realizes efficient and robust time series feature extraction through a three-level collaborative mechanism of causal convolution, dilated convolution and residual connection.

[0015] Further, in the process of predicting the outlet matte temperature, the key indicators are preliminarily predicted by the backbone network, and the predicted values and true values of the training set and test set are collected to construct a training sample set of model errors.

[0016] Further, in the process of predicting the outlet matte temperature, the first-order and second-order time difference methods are used to extract dynamic features to capture the instantaneous change rate and acceleration information of the data, and form a rich feature vector as the input of the XGBoost model.

[0017] Further, in the process of predicting the outlet matte temperature, the XGBoost model captures the nonlinear features of the error through iterative learning, and introduces a dynamic error compensation strategy based on local standard deviation to adjust the compensation weight according to the local fluctuation, thereby realizing differentiated processing of different fluctuation conditions.

[0018] Further, in the process of model optimization, the predicted values of the backbone model and the errors predicted by XGBoost are fused to obtain the final compensation prediction result.

[0019] The present application has the following beneficial effects:

[0020] The present application adopts the Pearson-Fscore fusion feature screening method, which can compress the input dimension under the premise of ensuring the integrity of the key information, and provide a high-quality, low-redundancy feature subset for subsequent modeling; the TCN algorithm model is used to extract global features, so that the model can capture long-distance dependencies without increasing the number of parameters; the backbone network is used to preliminarily predict the key indicators, and the predicted values and true values of the training set and test set are used to construct a training sample set of model errors. The first-order and second-order time difference methods are used to extract dynamic features, which are used as the input of the XGBoost model, further improving the ability to capture the nonlinear features of the error; a dynamic error compensation strategy based on local standard deviation is adopted, and the compensation weight is dynamically adjusted to adapt to different fluctuation conditions. By adaptively adjusting the size of the sliding window, this strategy improves the accuracy and stability of the model. Finally, the predicted values of the backbone model and the XGBoost error prediction are combined to obtain more accurate temperature prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a whole block diagram of matte temperature prediction of the present application.

[0022] Figure 2 It is a schematic diagram of the TCN of the present application.

[0023] Figure 3 The prediction result schematic diagram before and after error compensation of the application.

[0024] Figure 4 The prediction result schematic diagram of comparative experiments of STM, XGBoost, AGCRN and TCN-XGBoost of the application.

[0025] Figure 5 The interval prediction result schematic diagram of the application. DETAILED DESCRIPTION

[0026] The application will be further described below in combination with the drawings and specific embodiments.

[0027] A prediction method combining a time sequence convolution network and Xgboost dynamic error compensation, the overall block diagram of TCN-XGboost temperature prediction is as shown in the figure Figure 1 , which includes the following three parts:

[0028] 1. Pearson-Fscore fusion feature screening:

[0029] The DCS system of the nickel matte smelting production line monitors and records hundreds of process variables in real time, covering multiple subsystems such as charging, oxygen supply, dust removal, etc. For example, nickel concentrate quantity, compressed air main flow and temperature, oxygen main temperature and metering value, etc. High-dimensional and heterogeneous data contains rich information about the change of nickel matte temperature, but also brings feature redundancy, noise propagation, computational overhead and "dimension disaster" problems; a considerable part of the parameters has almost no significant effect on the temperature, and if all are directly input into the model, it will inevitably reduce the prediction accuracy and increase the training difficulty. Therefore, this paper adopts the Pearson-Fscore fusion de-redundancy (CF 3 R) feature screening strategy, which compresses the input dimension on the premise of ensuring the integrity of key information, and provides a high-quality, low-redundancy feature subset for subsequent modeling. The specific process is as follows:

[0030] Pearson correlation preliminary screening: for each candidate feature x i , calculate its Pearson correlation coefficient r i with the target variable (nickel matte temperature) y, considering that positive and negative correlations are equivalent in terms of prediction contribution, take the absolute value |r i | and arrange it in descending order, and keep the top 50% feature set S corr .

[0031] F-Score linear statistical screening: calculate F-Score (measure the linear explanation strength of the feature to y) for all features, and also select the top 50% feature set S F .

[0032] Comprehensive score: intersection S cand = S corr ∩ S F , for each feature in S cand , record its comprehensive score on Pearson correlation coefficient and F-Score respectively, calculate the ranking, and keep the top 25% of the set S core .

[0033] High correlation de-redundancy: calculate the correlation coefficient matrix R = [r core ] of each pair of features in S ij . If [r ij ] ≥ 0.95, it is determined to be highly collinear. According to the ri of each feature and the objective variable from small to large, if the current feature and the retained features are highly collinear, discard it, otherwise retain it. Get the de-redundant set S CFR .

[0034] Business key feature forced retention: based on metallurgical mechanism and production experience, for key variables such as "concentrate material quantity" S key , regardless of the screening result, it is retained. The final feature set is defined as S final = S CFR ∩ S key .

[0035] Through the above five steps, the original N-dimensional features are compressed to S final << N, which not only retains the information most explanatory to the temperature of nickel-molybdenum, but also significantly reduces the redundancy, providing high-quality input for model training.

[0036] 2. Convolutional network designed for time series data:

[0037] As shown in Figure 2 , TCN is a kind of convolutional network specially designed for time series data, unlike recurrent networks, TCN is composed entirely of convolutional layers, with one-dimensional causal convolution (Causal Conv) to ensure that the output at time t only depends on the input at time t and before, and with the help of residual connection (Residual Block) to stack deep network, taking into account long-term dependence modeling and training stability. The core idea is to use dilated convolution to quickly expand the receptive field, so that the model can capture long-distance dependencies without increasing the number of parameters, and the TCN structure is shown in Figure 2 .

[0038] Causal convolution: in time series modeling, it is necessary to ensure that the network can only access the input at time t and before when calculating the output at time t, and causal convolution is designed for this purpose:

[0039]

[0040] Compared with the cycle structure, it still maintains the advantage of one-time convolution parallel computing the whole sequence; compared with ordinary one-dimensional convolution, it can meet the strict time sequence causality.

[0041] Dilated convolution: In order to let the network see a longer history, directly increasing the convolution kernel or downsampling in the time domain will increase the parameters or lose the resolution, and the method of dilated convolution is to "stretch" the sampling interval in the convolution kernel to a given expansion rate d:

[0042]

[0043] When d = 1, it is equivalent to ordinary causal convolution; when d > 1, the input distance corresponding to adjacent convolution weights is enlarged to d time steps, and the receptive field is increased to (k-1)d+1.

[0044] Residual block: when deep dilated convolution layers are stacked, the gradient may be sharply attenuated or exploded in the back propagation process, and the training is easy to be unstable. To solve this problem, TCN adopts residual connection: add the original input x and the output F(x) obtained by convolution, if the input and output channel numbers are different, use a 1x1 convolution to do dimension alignment, get

[0045] y = F(x) + Conv 1×1 (x) #(3)

[0046] Let the network learn the identity mapping when needed, so that the gradient can pass through multiple layers without hindrance, while retaining the complex feature transformation provided by F(x), realizing deep and stable time series modeling.

[0047] 3、Xgboost error compensation:

[0048] Based on the backbone network, the key indicators of the nickel casting process are preliminarily predicted. By collecting the predicted value and the true value of the training set and the test set, a training sample set of model error is constructed. For the predicted results and the true value y i , the difference between and y i is extracted as the target label of the subsequent XGBoost error model, which provides the basis for subsequent compensation.

[0049] In the process of constructing the error compensation model, in order to deeply mine the dynamic characteristics of the nickel casting process data, this paper adopts the method of first-order and second-order time difference for feature extraction, and uses the extracted features as the input of the error compensation model. Since the data in the nickel casting process not only contains the state information of the current moment, but also contains the change trend of adjacent moments and across moments, the difference operation can obtain more rich dynamic characteristics. Let the original time series data be

[0050] X = x1, x2,..., x T #(1)

[0051] where x t denotes the data value at time t, and T is the length of the time series. The first-order difference Δx t can be represented as

[0052] Δx t = x t - x t-1 , t = 2,..., T##((2)

[0053] It can highlight the changes in data between adjacent time points, reflecting the instantaneous rate of change of data. It can reflect the instantaneous rate of change between adjacent time points. The second-order difference based on the first-order difference is further defined as

[0054] Δ 2 x t = Δx t - Δx t-1 = (x t - x t-1 ) - (x t-1 - x t-2 ) = x t - 2x t-1 + x t-2 , t = 3,..., T#(3)

[0055] The acceleration information for capturing data changes. Combining the first-order and second-order difference features, a more comprehensive and detailed feature vector can be constructed, providing more abundant and effective information for subsequent model training.

[0056] F = [Δx3, Δ 2 x3, Δx4, Δ 2 x4,..., Δx T , Δ 2 x T ]#(4)

[0057] Compared with traditional GBDT, XGBoost combines second-order derivative information and L1, L2 regularization terms in the objective function, greatly improving the accuracy of gradient estimation and effectively suppressing overfitting.

[0058] Through iterative learning of training set samples, XGBoost can capture the nonlinear characteristics of errors in each time period, providing accurate error prediction for subsequent compensation.

[0059] The optimization objective function of the model can be represented as:

[0060]

[0061] where is the loss function of the ith sample, is the predicted value, y i is the true value, Ω(x k ) is the regularization term of the kth tree. By performing a second-order Taylor expansion of the loss function, more accurate gradient information can be obtained:

[0062]

[0063] where

[0064]

[0065] To adapt to the large fluctuations or relatively stable situations that may occur in different time periods during the nickel casting process, a dynamic error compensation strategy based on local standard deviation is introduced into the model. By calculating the standard deviation sequence σ local (t) of the prediction error in the sliding window during the training phase and comparing it with the global standard deviation σ global (t), the global threshold threshold value

[0066] If the local fluctuation degree (σ local (t)) of a certain time period is significantly higher than the global threshold, it indicates that the time period is more intense, and the model will enable a higher compensation weight W high , otherwise the conservative correction coefficient W low is maintained. On this basis, if the local standard deviation is at a high level for a long time, in order to adapt to the actual situation of intense fluctuations, the sliding window size can be adaptively adjusted: when σ local (t) > 1.5σ global (t), by:

[0067] w t+1 = max(w t - δ, w min )#(1)

[0068] reduce the window length to better capture the current intense changes; conversely, when the fluctuations are stable, by:

[0069] w t+1 = min(w t + δ, w max )#(1)

[0070] increase the window to fully integrate historical information, thereby adaptively improving the robustness and precision of the model in dynamic scenarios.

[0071] The predicted value y base(t) the error e predicted by XGBoost xgb (t) the final compensated prediction result is obtained by fusion:

[0072] y final (t) = y base (t) + w(t) x e xgb (t) #(1)

[0073] where w(t) dynamically switches between {w high , w low} according to the local fluctuation, reflecting the differentiated processing of the dynamic compensation strategy for different fluctuations.

[0074] Limited by the inconsistent sampling period of nickel matte temperature label and process variable, in this paper, the same temperature represents the nickel matte temperature of the hour in each 60 min time window, so as to ensure the time alignment of the input and the label. In order to avoid the label overlap between the output samples, the window is sliding at 60 time steps, that is, a sample is generated every 60 minutes.

[0075] Due to the significant time autocorrelation of temperature evolution, the change of current nickel matte temperature y t is highly dependent on the temperature y t-1 ,..., y t-x of the previous period. In order to fully exploit this dynamic feature and improve the prediction accuracy, this paper introduces the temperature values of the previous period into the model input. Together with the process parameters {x t-x , …, x t-1 , x t} in the time period, it is used to predict y t . That is, the minute-level process information is retained, and the autoregressive term of nickel matte temperature is also introduced explicitly.

[0076] The data set used in this application has a total of 89941 minute-level original data, the input step is set to 180, the output step is set to 1, and the training set and test set are divided according to 8:2. After sample construction, the final number of training set samples is 1197, and the number of test set samples is 300.

[0077] In order to evaluate the performance of the proposed method, this paper uses the mean absolute error (MAE), the mean absolute percentage error (MAPE) and the correlation coefficient (R2) as the evaluation index of the model, and the formulas of the three indexes are shown in the formula.

[0078]

[0079] where y i represents the true value, represents the predicted value, N is the number of samples.

[0080] In the nickel smelting and casting process, the DCS system monitors hundreds of process variables, among which the pressure of the main nitrogen pipe and the total amount of circulating water have extremely limited impact on the nickel matte temperature. To eliminate redundant information and retain the features most relevant to temperature prediction, this paper uses a Pearson-Fscore fusion redundancy removal strategy to screen all variables. Ultimately, 10 key process variables were obtained as inputs to the nickel matte temperature prediction model, as shown in Table 1.

[0081] Table 1 Input Process Variables

[0082]

[0083] like Figure 3 As shown, a detailed comparative analysis of the predicted curves before and after compensation clearly reveals that in the 80–120, 180–220, and 240–260°C ranges, the peaks and valleys of the compensated curves more closely match the actual temperature curves. This phenomenon indicates that the differential feature plays a crucial role in capturing instantaneous deviations caused by sudden fluctuations during the casting process, effectively improving the model's response accuracy to dynamic changes. Furthermore, in the steady-state ranges of 0–60 and 140–170°C, the compensated curves almost completely overlap with the baseline, fully demonstrating that the framework optimizes the dynamic response without compromising the original steady-state accuracy, achieving a balance between dynamic and steady-state performance.

[0084] Table 2 Performance Indicators Before and After Error Compensation

[0085]

[0086] The performance metrics before and after interpolation are shown in Table 2. The XGBoost residual compensation strategy brought stable and significant performance gains to the TCN baseline model. The mean absolute error (MAE) decreased significantly from 7.6277 to 7.4941, a reduction of 1.75%, which means that the average deviation between the predicted and actual values ​​was further reduced; the root mean square error (RMSE) decreased from 9.7913 to 9.5732, a reduction of 2.23%, indicating that the mean square of the prediction error was significantly reduced and the overall accuracy of the model was improved; the coefficient of determination (R²) also decreased significantly. 2 The value was increased from 0.4415 to 0.4516.

[0087] The model proposed in this application has MAE 7.4941, RMSE 9.5732 and R 2 The optimal performance of 0.4516 confirms that it can more accurately suppress large errors, improve prediction accuracy and the ability to interpret temperature changes, and provide a more reliable solution for dynamic prediction of nickel matte exit temperature.

[0088] Table 3 Comparison of experimental performance indicators

[0089]

[0090] Observation of prediction results Figure 4 , LSTM (green curve) compared with the true value (blue curve), its fluctuations are chaotic and disorderly, and deviates from the true value in many areas, making it difficult to accurately capture the temperature change trend. XGBoost (orange curve) can follow the general trend of the true value in some sections, but there is still a significant deviation overall, and it cannot closely fit the fluctuations of the true temperature. AGCRN (red curve) has improved compared to the first two, but compared with the true value, there is still a certain degree of deviation, and it cannot finely match the real-time changes of the temperature. Looking at "TCN-XGBoost" (yellow curve), it is the closest to the blue true value curve, whether in the place where the temperature fluctuates sharply or in the relatively stable interval, it can more accurately echo the change trend of the true temperature, showing stronger capture ability and adaptability to the dynamic changes of the nickel-molybdenum outlet temperature, and directly presenting the superiority of the framework in predicting the nickel-molybdenum outlet temperature.

[0091] With the quantile regression ability of XGBoost, this paper trains three groups of error compensation models with τ = 0.05, 0.50, 0.95, and obtains a prediction interval covering about 90% probability. The results are shown in the figure, the blue star-shaped line represents the true temperature value, and the green area represents the 90% prediction interval. From Figure 5 it can be seen that the true value is covered by the green interval in most cases, even in areas where the temperature fluctuates sharply, such as the low point near sample index 100 and the peak near 200, the interval can still adapt to the changes and closely wrap the fluctuations of the true value. This shows that the prediction interval constructed by the three quantile models has good coverage and adaptability, and can provide reliable range estimation for the nickel-molybdenum temperature.

[0092] In view of the challenges of strong nonlinearity and large dynamic fluctuations of the process parameters, this study proposes a TCN and Xgboost dynamic error compensation fusion model: through Pearson-Fscore fusion to select key features such as furnace charging material components and air volume, use the expansion causal convolution of TCN to capture the dynamic correlation of multivariate time series, generate the initial prediction; introduce Xgboost to construct first-order / second-order time difference features, learn the error evolution law and realize dynamic compensation; based on Xgboost quantile regression to construct 90% prediction interval, quantify the prediction uncertainty.

[0093] Experiments show that compared with the baseline TCN model, the average absolute error (MAE) of the model decreases from 7.6277 to 7.4941 (decrease of 1.75%), the root mean square error (RMSE) decreases from 9.7913 to 9.5732 (decrease of 2.23%), the determination coefficient (R 2) to 0.4516, and significantly outperforms LSTM, Xgboost and other comparative models in both dynamic fluctuations (e.g., 80-120 sample segments) and steady-state intervals. The 90% prediction interval covers the true temperature well, especially in the case of severe fluctuations, and can still reliably contain the true value. Through the multi-level framework of "feature modeling-error compensation-interval estimation", this method effectively improves the prediction accuracy and robustness, providing an efficient solution for real-time prediction of nickel-molybdenum temperature and industrial process optimization.

Claims

1. A method for predicting the temperature of nickel matte based on temporal convolutional networks and dynamic error compensation, characterized by: include: (1) Feature selection method based on Pearson correlation coefficient and F-score: By using the fusion feature selection strategy of Pearson correlation coefficient (linear relationship) and F-score (non-linear relationship), the system compresses the input dimension while ensuring the integrity of key information, and provides a high-quality, low-redundancy feature subset for subsequent indicator prediction. (2) Based on the dilated causal convolution operation of the temporal convolutional network, the nonlinear coupling and temporal dynamic features between parameters are captured, thereby generating the initial prediction results. After the backbone network is used to make preliminary predictions on key indicators, the predicted values ​​and true values ​​of the training set and test set are collected, thereby constructing the training sample set of model error. (3) By introducing the initial prediction error into the Xgboost model based on first-order and second-order time difference features, the instantaneous rate of change and acceleration information of time series data are extracted through the training sample set, and the extracted features are used as the input of the error compensation model to learn the evolution law of the error and realize dynamic compensation. (4) Calculate the local standard deviation by sliding window, dynamically adjust the compensation weight and window length, distinguish between high / low fluctuation periods, increase the compensation intensity in the high fluctuation stage, and make conservative corrections in the low fluctuation stage to balance response speed and stability. (5) By fusing the prediction results of different models, the final compensated prediction result is obtained through prediction value fusion: y final (t)=y base (t)+w(t)×e xgb (t), thereby improving prediction accuracy and reducing prediction error, where w(t) depends on local fluctuations in {w high ,w low The dynamic switching between these states reflects the differentiated handling of different fluctuations in the dynamic compensation strategy.

Images