Time sequence prediction method and system for early warning and prediction of steel production cost

Abstract

The invention provides a time sequence prediction method for early warning and prediction of steel production cost, and the method comprises the steps: S1, collecting modeling data, aligning the modeling data, and obtaining a to-be-processed data set; s2, judging whether variables of the to-be-processed data set accord with a threshold range or not; if the result is yes, sorting according to time to obtain a training set and a test set, and further training a baseline model; if not, marking the data as abnormal data, and correcting the data; s3, training a model based on the training set to obtain a trained model; and S4, calculating data of the test set to obtain final prediction data. According to the method, with the help of a Transform architecture and a time sequence prediction model, and compared with a conventional time sequence model, such as a fixed model structure of ARIMA and LSTM, different modules are combined to adapt to diversified data characteristics, and a self-attention mechanism of the Transform architecture and a time sequence processing capability of a recurrent neural network are fused, so that the time sequence prediction is realized. And the prediction accuracy and stability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of data prediction technology, and specifically, relates to a time series prediction method and system for early warning prediction of steel production cost, and more specifically, to a time series prediction method and system for early warning prediction of steel production cost based on a large model. Background Art

[0002] Steel production cost early warning forecast refers to the forecast and analysis of the cost in the steel production process, so as to take timely measures to ensure that the cost is controlled within a reasonable range, thereby avoiding cost exceeding the budget or causing unnecessary waste. This process involves the investigation, analysis, and forecast of steel production costs, as well as the adoption of corresponding economic accounting and cost control measures based on the forecast results, aiming to improve corporate benefits and expand reproduction.

[0003] Specifically, the early warning forecast of steel production costs includes the following key aspects:

[0004] Cost forecasting: Through the analysis and forecasting of steel production costs, we can understand the cost change trend in the future and provide a basis for the company's cost control and decision-making. Early warning mechanism: Establish a cost early warning mechanism. When the cost approaches or exceeds the preset warning line, a warning signal will be issued in time to remind the company to take corresponding measures to make adjustments. Cost control: According to the forecast results, take effective cost control measures, such as optimizing production processes, reducing consumption, and improving production efficiency, to ensure that costs are controlled within the budget. The early warning forecasting system is crucial for steel companies because it can better manage costs and improve economic benefits. It also helps companies maintain their competitive advantage in the fierce market competition. Through scientific cost forecasting and analysis, unnecessary waste can be avoided, thereby achieving sustainable development.

[0005] At present, patent document CN116880191A discloses an intelligent control method for process industrial production systems based on time series prediction. The scheme relates to the field of process industrial technology. The method uses the historical time series data of any controlled variable and its affected manipulated variables and disturbance variables in the process industrial production system to train a time series prediction model for predicting the future value of the controlled variable, and then uses it as an evaluation equation to optimize the manipulated variables. Finally, the manipulated variables in the process industrial production system are optimized as a whole to obtain a unique target manipulated value for each manipulated variable so that the overall control process can automatically learn and fit the characteristics of nonlinear processes, and can simultaneously process multiple controlled variables and fit the characteristics of non-stationary processes, so that the overall control process is stable and controllable, and can solve the problems of external disturbances, multiple constraints, and complex mechanism modeling in the process of process industrial production, so as to achieve better control effects.

[0006] This patent provides an intelligent control method for process industrial production systems based on time series prediction. However, since it uses a fixed model structure, it cannot change the model structure according to actual production needs, increase or decrease specified modules, thereby reducing the generalization ability and scope of application of the model. Moreover, due to its single structure, it cannot handle the prediction needs of multiple time series, and it cannot utilize the relationship between different time series. Moreover, since the model does not use the baseline model as a reference for model prediction, it cannot accurately locate the problems in the model prediction process and make timely adjustments to the model. This problem needs to be solved urgently. Summary of the invention

[0007] In view of the defects in the prior art, the purpose of the present invention is to provide a time series prediction method and system for early warning prediction of steel production cost.

[0008] A time series prediction method for early warning and prediction of steel production cost provided by the present invention comprises:

[0009] Step S1: collecting modeling data, aligning the modeling data, and obtaining a data set to be processed;

[0010] Step S2: Determine whether the variables of the data set to be processed meet the threshold range; if the result is yes, sort them by time to obtain a training set and a test set; if the result is no, mark them as abnormal data and make corrections;

[0011] Step S3: training the model based on the training set to obtain a trained model;

[0012] Step S4: Calculate the data of the test set through the trained model to obtain the final prediction data.

[0013] Preferably, the model includes a baseline model and a complete model; the complete model uses the entire data as a time interval; the baseline model uses partial data as a time interval;

[0014] In the step S1, the modeling data is collected according to cost attribution; the cost attribution includes: process cost, energy consumption cost, other costs, steelmaking main raw material consumption cost and steelmaking auxiliary material consumption cost;

[0015] Aligning the modeling data, including:

[0016] Step A1: collecting the modeling data and dividing the time periods into days;

[0017] Step A2: Determine whether the acquisition frequency of the variables of the modeling data is greater than once a day; if the result is yes, execute step A3; if the result is no, execute step A4;

[0018] Step A3: Set weights and obtain the mean value of the variable on the day based on the weights, and record it in the data set to be processed;

[0019] Step A4: Calculate the characteristic value of the variable on that day by linear interpolation and record it in the data set to be processed.

[0020] Preferably, in step S2, the threshold range is 3% to 97% of the quantile of the variable on that day;

[0021] The variables of the data set to be processed are sorted by time, the first 75% of the variables are added as a training set, and the last 25% of the variables are added as a test set.

[0022] Preferably, in step S3, the preparatory processing of the training model includes:

[0023] Step B1: Based on the training set, encode the input data to obtain time series coding information;

[0024] Step B2: obtaining the weight of the time series according to the time series coding information;

[0025] Step B3: generating time characteristics of corresponding variables based on the weight of the time series;

[0026] Step B4: Based on the data of the training set, construct the mean square error as the objective function and train the model.

[0027] Preferably, in step S1, all variables of the data set to be processed are labeled with categorical / continuous, static / time-varying labels according to their properties; and according to the labels, each variable is dynamically scaled in the time series through different normalizers;

[0028] In step S4, based on the data of the test set, the actual prediction result is obtained by minimizing the quantile loss function, and the time series prediction interval is output through quantile regression to estimate the confidence interval;

[0029] The confidence interval estimate includes:

[0030] Step 1: Construct a quantile set;

[0031] Step 2: Based on the quantile set and the mean square error obtained in step S3, construct a quantile loss function set; in the quantile loss function set, each quantile loss function corresponds to a quantile in the quantile set;

[0032] Step 3: Minimize the quantile loss function of the quantile loss function set through the Adam algorithm to obtain the quantile prediction data under the corresponding quantile, evaluate and correct it in combination with the actual prediction result to obtain the final prediction result.

[0033] A time series prediction system for early warning and prediction of steel production cost provided by the present invention comprises:

[0034] Module M1: collect modeling data, align the modeling data, and obtain a data set to be processed;

[0035] Module M2: Determine whether the variables of the data set to be processed meet the threshold range; if the result is yes, sort them by time to obtain the training set and the test set, and then train the baseline model; if the result is no, mark them as abnormal data and make corrections;

[0036] Module M3: train the model based on the training set to obtain the trained model;

[0037] Module M4: Calculate the data of the test set through the trained model to obtain the final prediction data.

[0038] Preferably, the model includes a baseline model and a complete model; the complete model uses the entire data as a time interval; the baseline model uses partial data as a time interval;

[0039] In the module M1, the modeling data is collected according to cost attribution; the cost attribution includes: process cost, energy consumption cost, other costs, steelmaking main raw material consumption cost and steelmaking auxiliary material consumption cost;

[0040] Aligning the modeling data, including:

[0041] Module A1: Collect the modeling data and divide the time periods into days;

[0042] Module A2: Determine whether the acquisition frequency of the variables of the modeling data is greater than once a day; if the result is yes, trigger module A3; if the result is no, trigger module A4;

[0043] Module A3: Set weights and obtain the mean value of the variable on the day based on the weights, and record it in the data set to be processed;

[0044] Module A4: Calculate the characteristic value of the variable on that day by linear interpolation and record it in the data set to be processed.

[0045] Preferably, in the module M2, the threshold range is 3% to 97% of the quantile of the variable on that day;

[0046] The variables of the data set to be processed are sorted by time, the first 75% of the variables are added as a training set, and the last 25% of the variables are added as a test set.

[0047] Preferably, in the module M3, the preparatory processing of the training model includes:

[0048] Module B1: Based on the training set, encode the input data to obtain time series coding information;

[0049] Module B2: obtaining the weight of the time series according to the time series coding information;

[0050] Module B3: generating time characteristics of corresponding variables based on the weight of the time series;

[0051] Module B4: Based on the data of the training set, construct the mean square error as the objective function and train the model.

[0052] Preferably, in the module M1, all variables of the data set to be processed are labeled with categorical / continuous, static / time-varying labels according to their properties; and according to the labels, each variable is dynamically scaled in the time series through different normalizers;

[0053] In the module M4, based on the data of the test set, the actual prediction result is obtained by minimizing the quantile loss function, and the time series prediction interval is output through quantile regression to estimate the confidence interval;

[0054] The confidence interval estimate includes:

[0055] Module 1: Constructing quantile sets;

[0056] Module 2: constructing a quantile loss function set based on the quantile set and the mean square error obtained by module M3; in the quantile loss function set, each quantile loss function corresponds to a quantile in the quantile set;

[0057] Module 3: Minimize the quantile loss function of the quantile loss function set through the Adam algorithm to obtain the quantile prediction data under the corresponding quantile, evaluate and correct it in combination with the actual prediction result to obtain the final prediction result.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. The present invention introduces a time series prediction model based on a large model to predict the changing trend of steelmaking costs in the future, and conducts root cause analysis on the anomalies therein, and intuitively understands the influence of various factors on the steelmaking production cost by visualizing the attention weight.

[0060] 2. The present invention uses the Transformer architecture and time series prediction model. Unlike traditional time series models, such as ARIMA and LSTM, which have fixed model structures, the present invention adapts to diverse data characteristics by combining different modules, and integrates the self-attention mechanism of the Transformer architecture and the time series processing capability of the recurrent neural network, thereby improving the prediction accuracy and stability of the model.

[0061] 3. The present invention provides an estimate of the prediction interval through quantile prediction, helping production-related personnel understand the future distribution of variables. At the same time, combined with the visual attention mechanism, it allows users to more intuitively understand the direct and indirect factors of cost anomalies in the production process, making the output of the model more valuable for reference. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0063] Figure 1 A schematic diagram of the process provided by the present invention. DETAILED DESCRIPTION

[0064] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0065] The present invention introduces the large model architecture Transformer into the time series model, collects historical data on the production costs of each process, main raw material / auxiliary material consumption costs, energy consumption costs, other costs, etc. in the steel production process, and builds a comprehensive cost early warning and prediction system for the entire process. In addition to being able to predict the future consumption cost trend of each link based on the production data of each department, the system can also locate the cause of cost anomalies and issue early warnings to achieve the purpose of controlling production costs.

[0066] The innovation of this system is that it uses the Transformer architecture to divide the production data and cost data of each link into multiple time series with different distributions according to their correlation, and then trains the heterogeneous time series model by focusing on local processing of specific time series features and global processing that captures the collective features of all time series. At the same time, the system introduces multi-level prediction, and uses quantile regression to output the prediction interval, combined with the self-attention mechanism of the Transformer architecture, to derive the contribution of each parameter, so as to more accurately locate the cause of cost anomalies.

[0067] A time series prediction method for early warning and prediction of steel production cost provided by the present invention comprises:

[0068] Step 1: Data collection, that is, collecting all relevant data of the model according to the cost attribution of its impact;

[0069] In other words, the modeling data is collected and aligned to obtain a data set to be processed; the collected modeling data, that is, all data related to the model, is collected according to the attribution of its impact cost.

[0070] The costs include: process costs, energy consumption costs, other costs and steelmaking main raw material / auxiliary material consumption costs;

[0071] The process costs include: manufacturing costs, maintenance and repair costs, solvent consumption and material loss;

[0072] The main raw material / auxiliary material consumption cost of steelmaking includes: steel material consumption, scrap steel recycling, blast furnace auxiliary material consumption and converter auxiliary material consumption;

[0073] The energy consumption costs include: coke / coal powder consumption, industrial gas consumption, industrial oxygen / water consumption, and electricity consumption;

[0074] Other costs include: wages and benefits, fixed start-up costs, external warehousing costs, logistics costs;

[0075] The other costs mentioned also include other factors closely related to steelmaking production, including: steel market prices, monthly production plans, and changing trends in demand from downstream companies.

[0076] After collecting relevant data, data alignment is performed according to statistical methods. The alignment logic is:

[0077] 1. Collect statistical data of all relevant variables in the past 10 years and divide them into 3652 time periods with daily intervals;

[0078] 2. For variables with a collection frequency of more than once a day, set a unified time, set a weight based on the time difference between each collection time point and the time, and obtain the average value of the variable for the day through weighted averaging; the unified time is 12:00.

[0079] 3. For variables with a collection frequency of less than once a day, the characteristic value of the variable on that day is calculated using linear interpolation based on the collection frequency and the previous and subsequent data.

[0080] Step 2: Data preprocessing, creating datasets and baseline models;

[0081] Specifically, the collected data is arranged by time, and the normal threshold range of the statistical variables is followed according to the historical data. If there is data outside the 3% to 97% percentile on that day, the data is marked as abnormal data. An interface is added to the input end of the model to allow relevant business personnel to manually modify and fill in empty values ​​or abnormal value data. Finally, a time index is added according to the date to facilitate subsequent modeling work.

[0082] After completing the data preprocessing, all variables are labeled as categorical / continuous, static / time-varying according to their nature. For variables with different labels, different normalizers are used to dynamically scale each variable in the time series according to their distribution characteristics to reduce the forward-looking bias caused by normalization.

[0083] During training, the data is sorted by time, with the first 75% divided into the training set and the last 25% divided into the test set. A baseline model is also introduced, that is, the last year's data is trained separately, and the last 3 months of data is used as the test set of the baseline model. The purpose is to provide benchmark performance, simplify the problem, guide the final model selection, and quickly evaluate the model effect.

[0084] Step 3: Model initialization and training.

[0085] Specifically, first, in the initialization stage before training, key parameters such as the number of layers of the Transformer architecture, learning rate, embedding dimension, and number of heads of the attention mechanism are defined, and the exit probability is set to prevent the model from overfitting.

[0086] The model includes a baseline model and a complete model; the complete model uses the entire data as a time interval; the baseline model uses part of the data as a time interval;

[0087] Specifically, some of the data are from the most recent year;

[0088] In other words, in this embodiment, the model includes a complete model and a baseline model; the model with the overall data as the time interval is the complete model; the model with the last year's data as the time interval is the baseline model.

[0089] In particular, for static variables of the model input, their temporal features are enhanced by static covariates, and then a separate encoder-decoder and attention mechanism are used for each variable to determine the input contribution that changes over time. The input contribution is used to confirm the attention weight of the variable.

[0090] During the training process, the MSE, i.e., mean square error function, of the predicted value and the actual value is used to optimize the parameters of the model. By continuously adjusting the parameters to better predict the data, the best model is finally trained based on the training data set, i.e., the training set.

[0091] Specifically, the predicted value corresponds to the value obtained by the model during the training process on the training set, which refers to the predicted link cost; the actual value represents the actual value of the data on the training set, which refers to the actual link cost.

[0092] Step 4: Model evaluation and interval forecasting.

[0093] After the training is completed, according to the cost of each link divided in step 1, the data sampled based on the model on the test set is calculated respectively, and the prediction error between the prediction result and the true value on the actual test set is calculated to evaluate the prediction accuracy of the model.

[0094] In addition to the actual prediction results obtained by minimizing the quantile loss function, the present invention uses quantile regression to output the time series prediction interval, and estimates the confidence interval of the cost of each link at a specific time in the future, thereby improving the model prediction effect.

[0095] The steps for estimating confidence intervals using quantile regression are as follows:

[0096] Step 1: construct a set of output quantiles; the quantile set is determined according to actual needs, and the default value is {0.1, 0.5, 0.9};

[0097] Then, based on the MSE obtained in step 3, a weighted quantile loss function set containing the training set data is constructed; each quantile loss function in the quantile loss function set corresponds to a quantile in the quantile set; the loss function is the mean square error function, MSE;

[0098] Step 2: Add regularization terms to all quantile loss functions in the loss function set; the purpose of this step is to improve the efficiency of the loss function minimization process, improve the generalization ability of the model, and prevent the model from overfitting.

[0099] Step 3: For each quantile loss function, use the Adam adaptive learning rate algorithm to minimize it, and finally get the prediction result under the specific quantile;

[0100] In the step 3, in particular, when the quantile is 0.5, the quantile loss function degenerates into the MSE of step 3; this situation can be regarded as a special case of quantile regression.

[0101] Step 5: Model interpretation and visualization of attention weights;

[0102] This solution introduces the model explanation part, which analyzes the global time relationship of variables through the contribution of each variable parameter, and explains the global behavior of the model on the entire data set, so as to identify specific patterns such as persistent seasons or lag effects, and also to identify the environment in which certain phenomena occur. In addition, by analyzing the attention weight, the model can intuitively display the model's attention weight for different features and time points during prediction, and observe whether each time series has undergone significant changes due to outliers, so as to more quickly locate the variables directly affecting the cost anomaly.

[0103] Step 6: System deployment and self-learning. After the evaluation is completed, the model will be deployed to the production environment, connected to the production systems of each link and the system data of other relevant departments, and timely forecast and warn when there is a significant fluctuation in the interval prediction results. Regularly update and predict based on the baseline model generated in step 2, continue to track the performance of the model in actual applications, regularly evaluate the model prediction effect and readjust the model parameters as needed. If the model prediction deviation is large, the data should be re-collected and the self-learning process should be started, that is, execute steps 1 to 4.

[0104] The model architecture part of step 3 includes the following steps:

[0105] Step 3.1: Input encoding layer. The model encodes the input data, converts the categorical features into fixed-length vectors through the embedding layer, and directly retains the numerical form of the continuous features. At the same time, a position code is added to each time step to reflect the order information of the time series. Finally, the labels in step 2 are fused through a specific attention mechanism to enhance the model's sensitivity to data features.

[0106] Step 3.2: GRN unit, i.e. gated recurrent network and VSN, i.e. variable selection network.

[0107] The gated recurrent network refers to the incomplete data / discontinuous data. In order to prevent the gradient vanishing problem from inhibiting the network training, ELU, i.e., exponential linear unit and GLU, i.e., gated linear unit are used to screen the variables, and then residual connections are introduced to allow some layers to be skipped to accelerate training to alleviate the gradient vanishing problem.

[0108] The variable selection network refers to a variable weight assignment mechanism, that is, all past moments at a certain moment generate a normalized weight vector through a gated recurrent network unit and a softmax classification function. In addition, each variable is processed by a separate GRN unit to create a processing vector. Finally, the output of the VSN, that is, the variable selection network, is generated by a linear combination of the weight vector and the processing vector.

[0109] Step 3.3: Multi-head attention mechanism;

[0110] The temporal encoding information of the previous step is used through a multi-head attention mechanism to capture the long-term dependencies in the time series data. In each self-attention layer, the multi-head attention mechanism divides the input data into multiple parts, each of which is independently calculated by a "head". Each "head" learns different feature relationships based on different aspects of the input data, and adds a new matrix using seasonal analysis to enable different "heads" to share some weights, and then merges these results in the next step to form the final prediction result.

[0111] Step 3.4: Construct a temporal fusion decoder;

[0112] The Seq2Seq architecture is used to generate time features for each variable, and then the static covariates are used to enhance the time features in the static enrichment layer. The multi-head attention mechanism in step 3.3 is used in the time attention layer to obtain the weight of the variable time series, and it is ensured that each time dimension can only pay attention to the time series features before it in this process. Finally, a nonlinear GRN, namely a gene regulatory network and a residual connection are used in the position feedforward layer to speed up the training convergence.

[0113] Step 3.5: Model training;

[0114] After the above architecture is completed, based on the training set data, an optimization problem is constructed with MSE, or mean square error, as the objective function and the undetermined parameters in the model as the control variables. The gradient descent algorithm is used to find the optimal solution and assign it to the corresponding undetermined parameters.

[0115] The present invention also provides a time series prediction system for early warning and prediction of steel production costs. The time series prediction system for early warning and prediction of steel production costs can be realized by executing the process steps of the time series prediction method for early warning and prediction of steel production costs, that is, technical personnel in this field can understand the time series prediction method for early warning and prediction of steel production costs as a preferred implementation mode of the time series prediction system for early warning and prediction of steel production costs.

[0116] A time series prediction system for early warning and prediction of steel production cost provided by the present invention comprises:

[0117] Module M1: collect modeling data, align the modeling data, and obtain a data set to be processed;

[0118] Module M2: Determine whether the variables of the data set to be processed meet the threshold range; if the result is yes, sort them by time to obtain the training set and the test set, and then train the baseline model; if the result is no, mark them as abnormal data and make corrections;

[0119] Module M3: train the model based on the training set to obtain the trained model;

[0120] Module M4: Calculate the data of the test set through the trained model to obtain predicted data.

[0121] Specifically, in the module M1, the modeling data is collected according to cost attribution; the cost attribution includes: process cost, energy consumption cost, other costs, steelmaking main raw material consumption cost and steelmaking auxiliary material consumption cost;

[0122] Aligning the modeling data, including:

[0123] Module A1: Collect the modeling data and divide the time periods into days;

[0124] Module A2: Determine whether the acquisition frequency of the variables of the modeling data is greater than once a day; if the result is yes, trigger module A3; if the result is no, trigger module A4;

[0125] Module A3: Set weights and obtain the mean value of the variable on the day based on the weights, and record it in the data set to be processed;

[0126] Module A4: Calculate the characteristic value of the variable on that day by linear interpolation and record it in the data set to be processed.

[0127] Preferably, in the module M2, the threshold range is 3% to 97% of the quantile of the variable on that day;

[0128] The variables of the data set to be processed are sorted by time, the first 75% of the variables are added as a training set, and the last 25% of the variables are added as a test set.

[0129] Specifically, in the module M3, the preparatory processing of the training model includes:

[0130] Module B1: Based on the training set, encode the input data to obtain time series coding information;

[0131] Module B2: obtaining the weight of the time series according to the time series coding information;

[0132] Module B3: generating time characteristics of corresponding variables based on the weight of the time series;

[0133] Module B4: Based on the data of the training set, construct the mean square error as the objective function and train the model.

[0134] Specifically, in the module M1, all variables of the data set to be processed are labeled with categorical / continuous, static / time-varying labels according to their properties; and according to the labels, each variable is dynamically scaled in the time series through different normalizers.

[0135] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0136] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A time series prediction method for early warning prediction of steel production cost, characterized in that: include: Step S1: collecting modeling data, aligning the modeling data, and obtaining a data set to be processed; Step S2: Determine whether the variables of the data set to be processed meet the threshold range; if the result is yes, sort them by time to obtain a training set and a test set; if the result is no, mark them as abnormal data and make corrections; Step S3: training the model based on the training set to obtain a trained model; Step S4: Calculate the data of the test set through the trained model to obtain the final prediction data.

2. The time series prediction method for early warning and prediction of steel production cost according to claim 1 is characterized in that: The model includes a baseline model and a complete model; the complete model uses the entire data as a time interval; the baseline model uses part of the data as a time interval; In the step S1, the modeling data is collected according to cost attribution; the cost attribution includes: process cost, energy consumption cost, other costs, steelmaking main raw material consumption cost and steelmaking auxiliary material consumption cost; Aligning the modeling data, including: Step A1: collecting the modeling data and dividing the time periods into days; Step A2: Determine whether the acquisition frequency of the variables of the modeling data is greater than once a day; if the result is yes, execute step A3; if the result is no, execute step A4; Step A3: Set weights and obtain the mean value of the variable on the day based on the weights, and record it in the data set to be processed; Step A4: Calculate the characteristic value of the variable on that day by linear interpolation and record it in the data set to be processed.

3. The time series prediction method for early warning and prediction of steel production cost according to claim 2 is characterized in that: In step S2, the threshold range is 3% to 97% of the quantile of the variable on that day; The variables of the data set to be processed are sorted by time, the first 75% of the variables are added as a training set, and the last 25% of the variables are added as a test set.

4. The time series prediction method for early warning and prediction of steel production cost according to claim 3 is characterized in that: In step S3, the preparatory processing of the training model includes: Step B1: Based on the training set, encode the input data to obtain time series coding information; Step B2: obtaining the weight of the time series according to the time series coding information; Step B3: generating time characteristics of corresponding variables based on the weight of the time series; Step B4: Based on the data of the training set, construct the mean square error as the objective function and train the model.

5. The time series prediction method for early warning and prediction of steel production cost according to claim 4 is characterized in that: In the step S1, all variables of the data set to be processed are labeled with categorical / continuous, static / time-varying labels according to their properties; and according to the labels, each variable is dynamically scaled in the time series through different normalizers; In step S4, based on the data of the test set, the actual prediction result is obtained by minimizing the quantile loss function, and the time series prediction interval is output through quantile regression to estimate the confidence interval; The confidence interval estimate includes: Step 1: Construct a quantile set; Step 2: Based on the quantile set and the mean square error obtained in step S3, construct a quantile loss function set; in the quantile loss function set, each quantile loss function corresponds to a quantile in the quantile set; Step 3: Minimize the quantile loss function of the quantile loss function set through the Adam algorithm to obtain the quantile prediction data under the corresponding quantile, evaluate and correct it in combination with the actual prediction result to obtain the final prediction result.

6. A time series prediction system for early warning and prediction of steel production costs, characterized in that: include: Module M1: collect modeling data, align the modeling data, and obtain a data set to be processed; Module M2: Determine whether the variables of the data set to be processed meet the threshold range; if the result is yes, sort them by time to obtain the training set and the test set, and then train the baseline model; if the result is no, mark them as abnormal data and make corrections; Module M3: train the model based on the training set to obtain the trained model; Module M4: Calculate the data of the test set through the trained model to obtain the final prediction data.

7. The time series prediction system for early warning and prediction of steel production cost according to claim 6 is characterized in that: The model includes a baseline model and a complete model; the complete model uses the entire data as a time interval; the baseline model uses part of the data as a time interval; In the module M1, the modeling data is collected according to cost attribution; the cost attribution includes: process cost, energy consumption cost, other costs, steelmaking main raw material consumption cost and steelmaking auxiliary material consumption cost; Aligning the modeling data, including: Module A1: Collect the modeling data and divide the time periods into days; Module A2: Determine whether the acquisition frequency of the variables of the modeling data is greater than once a day; if the result is yes, trigger module A3; if the result is no, trigger module A4; Module A3: Set weights and obtain the mean value of the variable on the day based on the weights, and record it in the data set to be processed; Module A4: Calculate the characteristic value of the variable on that day by linear interpolation and record it in the data set to be processed.

8. The time series prediction system for early warning and prediction of steel production cost according to claim 7, characterized in that: In the module M2, the threshold range is 3% to 97% of the quantile of the variable on that day; The variables of the data set to be processed are sorted by time, the first 75% of the variables are added as a training set, and the last 25% of the variables are added as a test set.

9. The time series prediction system for early warning and prediction of steel production cost according to claim 8, characterized in that: In the module M3, the preparatory processing of the training model includes: Module B1: Based on the training set, encode the input data to obtain time series coding information; Module B2: obtaining the weight of the time series according to the time series coding information; Module B3: generating time characteristics of corresponding variables based on the weight of the time series; Module B4: Based on the data of the training set, construct the mean square error as the objective function and train the model.

10. The time series prediction system for early warning and prediction of steel production cost according to claim 9, characterized in that: In the module M1, all variables of the data set to be processed are labeled with categorical / continuous, static / time-varying labels according to their properties; and according to the labels, each variable is dynamically scaled in the time series through different normalizers; In the module M4, based on the data of the test set, the actual prediction result is obtained by minimizing the quantile loss function, and the time series prediction interval is output through quantile regression to estimate the confidence interval; The confidence interval estimate includes: Module 1: Constructing quantile sets; Module 2: constructing a quantile loss function set based on the quantile set and the mean square error obtained by module M3; in the quantile loss function set, each quantile loss function corresponds to a quantile in the quantile set; Module 3: Minimize the quantile loss function of the quantile loss function set through the Adam algorithm to obtain the quantile prediction data under the corresponding quantile, evaluate and correct it in combination with the actual prediction result to obtain the final prediction result.

Images