A method, system, device, and medium for temperature prediction of a hot blast stove dome

Abstract

The application provides a hot blast stove dome temperature prediction method, system, device and medium, relates to the hot blast stove temperature prediction technical field, and includes: obtaining the operation parameter data of the hot blast stove at the current time step; temperature feature matrix is constructed based on the operation parameter data, the temperature feature matrix includes the operation parameter data and the air-fuel ratio trend feature, the dome temperature multiscale feature, the waste gas temperature multiscale feature determined based on the operation parameter data; through the temperature prediction model obtained by pre-training, the context vector corresponding to the multi-interval of the temperature feature matrix is extracted based on the pre-trained multi-granularity spatiotemporal attention mechanism, and the context vector of the multi-interval is linearly combined through the gate weight, so as to predict the dome temperature prediction value and the waste gas temperature prediction value of the hot blast stove at the future time step. The application realizes accurate prediction of the hot blast stove dome temperature and the waste gas temperature.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hot blast stove temperature prediction, in particular to a hot blast stove dome temperature prediction method, system, device and medium. BACKGROUND

[0002] The hot blast stove is a kind of high-temperature furnace, which is widely used in metallurgical industry. During the operation of the hot blast stove, accurate prediction of the dome temperature and the exhaust gas temperature is crucial, as it directly affects the safety of production. Traditional temperature control methods rely on real-time detection of the dome temperature and the exhaust gas temperature by sensors, and then control and adjust the hot blast stove, which is difficult to adapt to rapidly changing production demands.

[0003] In recent years, with the development of artificial intelligence technology, data-driven temperature prediction methods have been researched and developed. These methods use historical operation data to predict future temperatures to achieve more accurate temperature control. Although these methods have made some progress, there are still some limitations. For example, using only basic time series features ignores other feature representations, mixing short-term and long-term trend features in the hidden layer hinders future temperature prediction, and single-step prediction mode leads to error accumulation. SUMMARY

[0004] Therefore, the present application aims to provide a hot blast stove dome temperature prediction method, system, device and medium, which realizes accurate prediction of the hot blast stove dome temperature and the exhaust gas temperature.

[0005] In a first aspect, the present application provides a hot blast stove dome temperature prediction method, comprising:

[0006] Obtaining the operation parameter data of the hot blast stove at the current time step, the operation parameter data including dome temperature time series data, exhaust gas temperature time series data, air-fuel ratio time series data, air flow time series data and coal gas flow time series data;

[0007] Determining the air-fuel ratio trend feature based on the air-fuel ratio time series data in the operation parameter data, and determining the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature based on the dome temperature time series data and the exhaust gas temperature time series data in the operation parameter data, respectively, and constructing a temperature feature matrix combining the operation parameter data, the air-fuel ratio trend feature, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature;

[0008] Using the pre-trained temperature prediction model, extracting the context vectors of multiple intervals corresponding to the temperature feature matrix based on the pre-trained multi-granularity spatio-temporal attention mechanism, and performing linear combination of the gate weights of the context vectors of multiple intervals to predict the dome temperature prediction value and the exhaust gas temperature prediction value of the hot blast stove at the future time step.

[0009] In an embodiment, a trend feature of the air-fuel ratio is determined based on the air-fuel ratio time series data in the operation parameter data, and a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature are respectively determined based on the dome temperature time series data and the exhaust gas temperature time series data in the operation parameter data, and a temperature feature matrix is constructed by combining the operation parameter data, the trend feature of the air-fuel ratio, the dome temperature multi-scale feature, and the exhaust gas temperature multi-scale feature, including:

[0010] For the air-fuel ratio at any time except the first time in the air-fuel ratio time series data, a neighboring air-fuel ratio difference feature between the air-fuel ratio at the current time and the air-fuel ratio at the previous time is determined based on a preset air-fuel ratio reference value, and an air-fuel ratio trend feature is constructed based on the neighboring air-fuel ratio difference features at all times;

[0011] For any time series data in the dome temperature time series data and the exhaust gas temperature time series data, a window ratio corresponding to the time series data at different scales is determined, the mean feature and the standard deviation feature corresponding to the time series data are determined according to the window ratio, and a multi-scale feature corresponding to the time series data is constructed, the multi-scale feature being divided into a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature;

[0012] The dome temperature time series data, the exhaust gas temperature time series data, the air-fuel ratio time series data, the air flow time series data, the gas flow time series data, the dome temperature multi-scale feature, and the exhaust gas temperature multi-scale feature are spliced to obtain a temperature feature matrix.

[0013] In an embodiment, the temperature prediction model includes an input layer, an improved LSTM layer, and a fully connected output layer; by using the temperature prediction model obtained by pre-training, a multi-interval context vector corresponding to the temperature feature matrix is extracted by using a pre-trained multi-granularity spatio-temporal attention mechanism, and the multi-interval context vector is linearly combined by a gating weight to predict a dome temperature prediction value and / or an exhaust gas temperature prediction value of the hot blast furnace at a future time step, including:

[0014] The temperature feature matrix is received by the input layer;

[0015] The short-term context vector, the medium-term context vector, and the long-term context vector are extracted from the temperature feature matrix by using the pre-trained multi-granularity spatio-temporal attention mechanism by the improved LSTM layer, and the short-term context vector, the medium-term context vector, and the long-term context vector are linearly combined by a gating weight to obtain a target context vector;

[0016] The dome temperature prediction value and / or the exhaust gas temperature prediction value of the hot blast furnace at the future time step are predicted based on the target context vector by the fully connected output layer.

[0017] In an embodiment, a pre-trained multi-granularity spatio-temporal attention mechanism is used to extract short-term context vectors, medium-term context vectors and long-term context vectors from the temperature feature matrix, and the target context vector is obtained by linearly combining the short-term context vectors, the medium-term context vectors and the long-term context vectors with gating weights, comprising:

[0018] According to the preset multi-interval ratio, the hidden features of the temperature feature matrix are divided into a short-term feature matrix, a medium-term feature matrix and a long-term feature matrix;

[0019] The region attention calculation is performed on the short-term feature matrix, the medium-term feature matrix and the long-term feature matrix respectively to obtain the short-term context vectors, the medium-term context vectors and the long-term context vectors;

[0020] The first weight matrix of the short-term context vectors, the medium-term context vectors and the long-term context vectors is determined respectively, and the target context vector is obtained by linearly combining the short-term context vectors, the medium-term context vectors and the long-term context vectors with gating weights.

[0021] In an embodiment, the region attention calculation is performed on the short-term feature matrix, the medium-term feature matrix and the long-term feature matrix respectively to obtain the short-term context vectors, the medium-term context vectors and the long-term context vectors, comprising:

[0022] The following operations are performed on any feature matrix in the short-term feature matrix, the medium-term feature matrix and the long-term feature matrix:

[0023] The hidden state of the feature matrix is nonlinearly mapped to obtain a mapping feature;

[0024] The second weight matrix of the corresponding hidden state of the mapping feature is determined based on the mapping feature, and the hidden state is weighted and aggregated to obtain the context vector corresponding to the feature matrix.

[0025] In an embodiment, after the region attention calculation is performed on the short-term feature matrix, the medium-term feature matrix and the long-term feature matrix respectively to obtain the short-term context vectors, the medium-term context vectors and the long-term context vectors, the method further comprises:

[0026] Based on the short-term context vectors, it is detected whether there is an abnormal event;

[0027] If yes, the authenticity of the abnormal event is verified based on the similarity between the short-term context vectors and the medium-term context vectors;

[0028] In the case that the abnormal event is a real abnormal event, the short-term context vectors are corrected, and a new long-term context vector is determined based on the corrected short-term context vectors and the long-term context vectors.

[0029] In an implementation, the method further comprises:

[0030] If the vault temperature prediction value and / or the exhaust gas temperature prediction value is higher than the dynamic early warning threshold, temperature warning information is generated.

[0031] In a second aspect, the present application further provides a temperature prediction device for a hot blast stove vault, comprising:

[0032] A data acquisition module is configured to acquire operation parameter data of the hot blast stove at a current time step, the operation parameter data comprising vault temperature time series data, exhaust gas temperature time series data, air-fuel ratio time series data, air flow time series data, and coal gas flow time series data.

[0033] A feature construction module is configured to determine an air-fuel ratio trend feature based on the air-fuel ratio time series data in the operation parameter data, and determine a vault temperature multi-scale feature and an exhaust gas temperature multi-scale feature based on the vault temperature time series data and the exhaust gas temperature time series data in the operation parameter data, respectively, and construct a temperature feature matrix by combining the operation parameter data, the air-fuel ratio trend feature, the vault temperature multi-scale feature, and the exhaust gas temperature multi-scale feature.

[0034] A temperature prediction module is configured to extract a multi-interval context vector corresponding to the temperature feature matrix based on a pre-trained multi-granularity spatio-temporal attention mechanism by using a pre-trained temperature prediction model, and perform a gate weight linear combination on the multi-interval context vector, so as to predict a vault temperature prediction value and / or an exhaust gas temperature prediction value of the hot blast stove at a future time step.

[0035] In a third aspect, the present application further provides an electronic device, comprising a processor and a memory, the memory storing computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the method of any one of the first aspect.

[0036] In a fourth aspect, the present application further provides a computer readable storage medium, the computer readable storage medium storing computer executable instructions, and the computer executable instructions, when invoked and executed by a processor, cause the processor to implement the method of any one of the first aspect.

[0037] The application provides a hot blast stove dome temperature prediction method, system, device and medium. First, operation parameter data of the hot blast stove at a current time step is acquired, and the operation parameter data includes dome temperature time series data, waste gas temperature time series data, air-fuel ratio time series data, air flow time series data and coal gas flow time series data. Then, a temperature feature matrix is constructed based on the operation parameter data, and the temperature feature matrix includes the operation parameter data and air-fuel ratio trend features, dome temperature multi-scale features and waste gas temperature multi-scale features determined based on the operation parameter data. Finally, a temperature prediction model obtained through pre-training is used to extract a multi-interval context vector corresponding to the temperature feature matrix based on a pre-trained multi-granularity spatio-temporal attention mechanism, and the multi-interval context vector is linearly combined with a gating weight to predict dome temperature prediction values and waste gas temperature prediction values of the hot blast stove at a future time step. The above method can realize accurate prediction of the dome temperature and the waste gas temperature of the hot blast stove at the future time step, significantly improve the reliability of temperature prediction, improve the combustion efficiency and reduce energy consumption by constructing the temperature feature matrix including the operation parameter data and the air-fuel ratio trend features, the dome temperature multi-scale features and the waste gas temperature multi-scale features determined based on the operation parameter data, and using the temperature prediction model to perform multi-granularity spatio-temporal attention processing on the temperature feature matrix to form the multi-interval context vector, and then combining the gating mechanism to linearly combine the multi-interval context vector with the gating weight.

[0038] Other features and advantages of the present application will be set forth in the descriptions below, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the description, claims and drawings.

[0039] In order to make the above-mentioned objects, characteristics and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0041] Figure 1 A flowchart of a hot blast stove dome temperature prediction method provided by an embodiment of the present application is shown in the figure;

[0042] Figure 2 A structure diagram of a hot blast stove dome temperature prediction method provided by an embodiment of the present application is shown in the figure;

[0043] Figure 3 A schematic diagram of an LSTM-MGSTA network provided for an embodiment of the present application;

[0044] Figure 4 A structural schematic diagram of a temperature prediction device for a hot blast stove dome provided for an embodiment of the present application;

[0045] Figure 5 A structural schematic diagram of an electronic device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in connection with the embodiments. Obviously, the described embodiments are only some, but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0047] In view of the poor coordination precision of the dome temperature and the exhaust gas temperature in the operation of the hot blast stove in the metallurgical industry, the present application provides a temperature prediction method, system, device and medium for a hot blast stove dome, which realizes accurate prediction of the dome temperature and the exhaust gas temperature of the hot blast stove. The embodiments of the present application solve the technical bottlenecks of the response delay of traditional methods in sudden temperature and the insufficient capture of long-term trend by combining a long short-term memory (LSTM) network with a multi-granularity spatio-temporal attention mechanism. The technical solution of the embodiments of the present application is suitable for temperature prediction of a hot blast stove, and provides core algorithm support for equipment safety warning.

[0048] To facilitate the understanding of the present embodiment, first, a temperature prediction method for a hot blast stove dome disclosed by the present embodiment will be described in detail, referring to a flowchart of a temperature prediction method for a hot blast stove dome shown in Figure 1 The method mainly includes the following steps S102 to S106:

[0049] Step S102, obtaining the operation parameter data of the hot blast stove at the current time step.

[0050] The operation parameter data includes dome temperature time series data, exhaust gas temperature time series data, air-fuel ratio time series data, air flow time series data and coal gas flow time series data.

[0051] In step S104, a trend feature of the air-fuel ratio is determined based on the air-fuel ratio time sequence data in the operation parameter data, and a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature are respectively determined based on the dome temperature time sequence data and the exhaust gas temperature time sequence data in the operation parameter data, and a temperature feature matrix is constructed by combining the operation parameter data, the trend feature of the air-fuel ratio, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature.

[0052] The temperature feature matrix includes the operation parameter data and the trend feature of the air-fuel ratio, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature determined based on the operation parameter data, the dome temperature multi-scale feature includes mean value features and standard deviation features of the dome temperature time sequence data at micro-scale, medium-scale and macro-scale, and the exhaust gas temperature multi-scale feature includes mean value features and standard deviation features of the exhaust gas temperature time sequence data at micro-scale, medium-scale and macro-scale.

[0053] In step S106, a multi-interval context vector corresponding to the temperature feature matrix is extracted based on the pre-trained multi-granularity spatio-temporal attention mechanism by using the temperature prediction model obtained by pre-training, and the multi-interval context vector is linearly combined by using a gating weight to predict a dome temperature prediction value and an exhaust gas temperature prediction value of the hot blast furnace at a future time step.

[0054] The multi-interval context vector includes a short-term context vector, a medium-term context vector and a long-term context vector. In an embodiment, an LSTM-MGSTA network is constructed as the temperature prediction model, an input layer receives the temperature feature matrix, a multi-granularity spatio-temporal attention mechanism is added after an improved LSTM layer in the model, M time step hidden layer vectors output by the improved LSTM layer are divided into three intervals of short-term, medium-term and long-term, a regional attention calculation is performed on each interval to obtain a regional context vector, and a final context is generated by linearly combining the regional context vectors by using a gating mechanism, and the finally generated context is input into a fully connected layer to predict dome temperatures and exhaust gas temperatures at N future time steps.

[0055] The temperature prediction method for the dome of the hot blast furnace provided by the embodiment of the present application can accurately predict dome temperatures and exhaust gas temperatures of the hot blast furnace at future time steps by constructing a temperature feature matrix including operation parameter data and a trend feature of an air-fuel ratio, a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature determined based on the operation parameter data, and performing multi-granularity spatio-temporal attention processing on the temperature feature matrix by using a temperature prediction model to form a multi-interval context vector, and then linearly combining the multi-interval context vector by using a gating mechanism, thereby significantly improving the reliability of temperature prediction, improving combustion efficiency and reducing energy consumption.

[0056] For the convenience of understanding, the embodiment of the present application provides a schematic diagram of the temperature prediction method for the dome of the hot blast furnace as shown in FIG. 4. Figure 2The application discloses a temperature prediction method for a hot blast stove dome, and relates to a training and prediction two-stage method.

[0057] The application takes the prediction stage as an example to explain the specific implementation process of the temperature prediction method for the hot blast stove dome.

[0058] (I) obtaining the running parameter data of the hot blast stove at a current time step, that is, collecting the running parameter data of the hot blast stove, including gas flow, air flow, dome temperature and waste gas temperature, and calculating the air-fuel ratio according to the following formula : , is the gas flow value, is the air flow value.

[0059] (II) constructing a temperature feature matrix based on the running parameter data. The feature tensor is constructed, which contains the following time step aligned features:

[0060] (2.1) basic time sequence features: dome temperature, waste gas temperature, air-fuel ratio, air flow and gas flow at the current time, forming an M*5 dimensional matrix.

[0061] (2.2) for the air-fuel ratio at any time except the first time in the air-fuel ratio time sequence data, based on the preset air-fuel ratio reference value, the adjacent air-fuel ratio difference feature between the air-fuel ratio at the current time and the air-fuel ratio at the previous time is determined, and the air-fuel ratio trend feature is constructed based on the adjacent air-fuel ratio difference features at all times. Specifically:

[0062] The adjacent air-fuel ratio difference features at the 2th to Mth time steps are calculated, the first time step is zero-filled, forming an M*1 dimensional matrix; the feature formula is:

[0063] ;

[0064] Where t=2,3,...,M is the tth time, is the adjacent air-fuel ratio difference feature at the tth time, is the air-fuel ratio at the tth time, is the air-fuel ratio at the tth time, is the air-fuel ratio at the tth time, is the air-fuel ratio according to the best combustion efficiency of the current working condition of the hot blast stove.

[0065] (2.3) For any one of the dome temperature time series data and the exhaust gas temperature time series data, determine the window proportion corresponding to the time series data at different scales, determine the mean feature and the standard deviation feature corresponding to the time series data according to the window proportion, and construct the multi-scale feature corresponding to the time series data, the multi-scale feature being divided into dome temperature multi-scale features and exhaust gas temperature multi-scale features.

[0066] The multi-scale feature is also a short name of multi-scale local binary pattern (MSD-LBP) feature: the mean and the standard deviation of the micro, medium and macro scales are calculated for the dome temperature and the exhaust gas temperature sequence, and then broadcast to all time steps to form an Mx12-dimensional matrix. Specifically, it includes:

[0067] A dynamic window proportion associated with the length of the time sequence M is used to avoid the limitations of a fixed window:

[0068] ;

[0069] wherein, for the scale , , , , , ,

[0070] The LBP features of the micro, medium and macro scales are calculated for the dome temperature and the exhaust gas temperature respectively, and the formula is as follows:

[0071] ;

[0072] wherein, is the LBP feature under the time window , , , , , , is the temperature value in the time window , , is a step function, is a variable. For example, assuming that the length of the time window is , the temperature value in the time window is , and the above formula can be obtained by .

[0073] The mean and the standard deviation of the LBP feature of each scale are extracted, that is, the LBP sequence of each scale is converted to , , that is, the mean, That is, the feature dimensions of the standard deviation, the dome temperature and the exhaust gas temperature are Mx6 dimensions, the MSD-LBP features of the dome temperature and the exhaust gas temperature are spliced to obtain a 12-dimensional feature vector, and the 12-dimensional feature vector is broadcast to M time steps to obtain an Mx12-dimensional feature vector.

[0074] (2.4) The dome temperature time series data, the exhaust gas temperature time series data, the air-fuel ratio time series data, the air flow time series data, the coal gas flow time series data, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature are spliced to obtain a temperature feature matrix, that is, the above features are spliced into an Mx18-dimensional feature matrix.

[0075] Further, the maximum and minimum value normalization processing can be performed on each feature column, and the normalization parameters are recorded to facilitate subsequent prediction result reverse normalization.

[0076] (Three) Through the temperature prediction model obtained by pre-training, the multi-granularity spatio-temporal attention mechanism is used to extract the context vectors of the temperature feature matrix corresponding to the multi-interval, and the context vectors of the multi-interval are linearly combined by the gating weight to predict the dome temperature prediction value and / or the exhaust gas temperature prediction value of the hot blast stove at the future time step.

[0077] First, the temperature prediction model is explained. The temperature prediction model includes an input layer, an improved LSTM layer and a fully connected output layer. Specifically, the input layer: receives a tensor with a shape of [M, 18]; the LSTM layer: uses the time series modeling capability of the LSTM network to extract long-term dependencies and deep time series features from the input features, and introduces a spatio-temporal attention mechanism after the LSTM layer. The multi-granularity spatio-temporal attention mechanism divides the hidden layer vectors of M time steps into three intervals of short-term, medium-term and long-term, performs feature projection, regional attention weight calculation and context vector generation on each interval, splices the context vectors of each interval, generates fusion weights through a trainable parameter matrix, and finally generates the final context by linear combination according to the gating weight. The fully connected layer output: the features processed by the multi-granularity spatio-temporal attention mechanism are output to the prediction value of the future N minutes of temperature through the fully connected layer.

[0078] The training process of the temperature prediction model is then explained: To improve the model's generalization ability, a K-fold crossover strategy is used to divide historical data into five equal-length data blocks based on temporal continuity, with each block containing a complete combustion cycle. In the solution space, the optimal hyperparameter combination, including the learning rate and the number of hidden layer units, is searched using a grid search method with the goal of minimizing the validation set MAE. The model is trained based on the optimal hyperparameter combination obtained from the search. During the training phase, the teacher-forced strategy inputs the true value into the network with probability p, and p linearly decays from 1.0 to 0.5 with the training epochs and the loss value. When the validation set MAE does not decrease within five consecutive epochs, the current model parameters are saved and training is terminated. The trained model structure and weights are saved to the file system for later retrieval. The saved content includes: the model structure, i.e., the configuration of the LSTM network and the multi-granularity spatiotemporal attention mechanism; weight parameters; and data normalization parameters, used for denormalization of predicted values.

[0079] Based on the trained temperature prediction model described above, real-time temperature prediction can be achieved, such as... Figure 3 The diagram shown illustrates an LSTM-MGSTA network, which specifically includes the following processes:

[0080] (3.1) The temperature feature matrix is ​​received through the input layer.

[0081] (3.2) By using the improved LSTM layer, the short-term context vector, medium-term context vector and long-term context vector are extracted from the temperature feature matrix using a pre-trained multi-granularity spatiotemporal attention mechanism. The short-term context vector, medium-term context vector and long-term context vector are then linearly combined with gated weights to obtain the target context vector.

[0082] In one example, the process of multi-granular spatiotemporal attention is as follows:

[0083] (a) According to the preset multi-interval ratio, the hidden features of the temperature feature matrix are divided into short-term feature matrix, medium-term feature matrix and long-term feature matrix. The M time steps are divided into three time intervals according to the ratio: long-term interval: the first 30% of time steps, i.e. Tlong=[1,0.3M]; medium-term interval: the middle 50% of time steps, i.e. Tmid=(0.3M,0.8M]; short-term interval: the last 20% of time steps, i.e. T short=(0.8M,M).

[0084] (b) Perform region attention calculations on the short-term, medium-term, and long-term feature matrices respectively to obtain the short-term, medium-term, and long-term context vectors. For any of the short-term, medium-term, and long-term feature matrices, perform the following operation:

[0085] (b1) Obtain the mapped features by performing a nonlinear mapping on the hidden states of the feature matrix. This is achieved by performing a nonlinear mapping on the hidden states using independent trainable linear transformation matrices. In one example, the calculation... ,in, For mapping features, For interval The projection matrix corresponding to 2, In hidden state, 2 represents the time interval Category 2∈{long,mid,short} Indicates time interval At any time within, This represents the hyperbolic tangent function.

[0086] (b2) Determine the second weight matrix of the hidden state based on the mapping features, and use it to perform weighted aggregation of the hidden states to obtain the context vector corresponding to the feature matrix. In one example, first, the mapping features are summed and normalized along the channel dimension, and then the second weight matrix of the hidden state is determined according to the following formula. : Then, the hidden states are weighted and aggregated to generate a context vector. : .in, Let represent the second weight matrix, indicating time . Relative to time interval The importance of For the mapping feature, the first One element, The total number of elements contained in the mapping feature. This is the context vector.

[0087] (c) Perform cross-scale optimization on short-term, medium-term, and long-term context vectors, specifically:

[0088] (c1) Detect the presence of abnormal events based on short-term context vectors.

[0089] Specifically, based on the hidden states of the short-term feature matrix and the second weight matrix, an anomaly index is defined: ;in, As an anomaly index, This is the second weight matrix of the short-term feature matrix. The historical mean of the second weight matrix of the short-term feature matrix. The hidden state of the short-term feature matrix. This represents the historical mean of the hidden states in the short-term feature matrix. (Judgment) , is a threshold value; if yes, it is determined that there is an abnormal event, otherwise there is no abnormal event.

[0090] (c2) if yes, the authenticity of the abnormal event is verified based on the similarity between the short-term context vector and the medium-term context vector.

[0091] Specifically, the cosine similarity between the short-term context vector and the medium-term context vector is calculated, and in the case that the cosine similarity is greater than a preset threshold, it is determined to be a real abnormal event, otherwise it is a false abnormal event.

[0092] (c3) in the case that the abnormal event is a real abnormal event, the short-term context vector is corrected, and a new long-term context vector is determined based on the corrected short-term context vector and the long-term context vector.

[0093] Specifically, if it is a false abnormal event, the second weight matrix of the short-term feature matrix is weakened. If it is a real abnormal event, the second weight matrix of the short-term feature matrix is enhanced to correct the short-term context vector. The corrected short-term context sequence is fused with the original long-term context vector to obtain a new long-term context vector.

[0094] (d) determine the first weight matrix of the short-term context vector, the medium-term context vector and the long-term context vector respectively, to linearly combine the short-term context vector, the medium-term context vector and the long-term context vector with the gate weight to obtain the target context vector.

[0095] Specifically, the process of the gating mechanism is as follows: the interval context vectors are spliced into , and then the fusion weight is generated through the trainable parameter matrix, where is the weight matrix, is the bias term. Finally, the target context vector is generated by linear combination with the gate weight , , , , , ,

[0096] (3.3) through the full connection output layer, based on the target context vector, the predicted value of the dome temperature and / or the predicted value of the exhaust gas temperature of the hot blast furnace at the future time step is predicted.

[0097] In an embodiment, if the vault temperature prediction value and / or the exhaust gas temperature prediction value is higher than the dynamic warning threshold, temperature warning information is generated. In practical applications, based on real-time data, a dynamic warning threshold is set, and when the prediction value exceeds the safety range, an alarm is automatically triggered.

[0098] In summary, the temperature prediction of the hot blast stove vault according to the embodiments of the present application aims to effectively prevent the hot blast stove from burning through by accurately predicting key temperature indicators. The embodiments of the present application not only solve the problem of insufficient single-scale modeling by using multi-scale joint representation features, but also solve the problem of insufficient model perception granularity and prediction accuracy by calculating and gate-fusing attention in each interval, thereby providing important technical support for the operation safety and energy utilization rate of hot blast stoves in industrial scenarios. Specifically, by fusing multi-scale features and a spatiotemporal attention mechanism and combining the time series modeling capability of an LSTM network, the embodiments of the present application achieve the following significant beneficial effects:

[0099] (1) The prediction accuracy is significantly improved: by introducing multi-scale features and a spatiotemporal attention mechanism, the dynamic change trend of the vault temperature and the exhaust gas temperature is modeled in depth, and the prediction accuracy is significantly improved.

[0100] (2) Efficient prediction method: a one-time prediction method is used to directly output the temperature sequence of the target time period, avoiding the error accumulation problem of rolling prediction, reducing the computational overhead, and improving the prediction efficiency.

[0101] (3) Robustness and generalization ability optimization: through K-fold cross-validation and teacher forced training strategies, the stability and generalization ability of the model under complex working conditions are optimized, so that it can adapt to different fuel heat value fluctuations, air pressure changes and other interference factors.

[0102] On the basis of the foregoing embodiments, the embodiments of the present application provide a temperature prediction device for a hot blast stove vault, as shown in Figure 4 The device mainly includes the following parts:

[0103] The data acquisition module 402 is configured to acquire the operation parameter data of the hot blast stove at the current time step, and the operation parameter data includes vault temperature time series data, exhaust gas temperature time series data, air-fuel ratio time series data, air flow time series data, and coal gas flow time series data.

[0104] The feature construction module 404 is configured to determine an air-fuel ratio trend feature based on the air-fuel ratio time sequence data in the operation parameter data, and determine a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature based on the dome temperature time sequence data and the exhaust gas temperature time sequence data in the operation parameter data respectively, and construct a temperature feature matrix by combining the operation parameter data, the air-fuel ratio trend feature, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature.

[0105] The temperature prediction module 406 is configured to extract a multi-interval context vector corresponding to the temperature feature matrix based on a pre-trained multi-granularity spatio-temporal attention mechanism of a temperature prediction model, and perform a gated weight linear combination on the multi-interval context vector to predict a dome temperature prediction value and / or an exhaust gas temperature prediction value of the hot blast stove at a future time step.

[0106] The temperature prediction device for the dome of the hot blast stove provided in the embodiment of the present application can construct a temperature feature matrix containing operation parameter data and an air-fuel ratio trend feature, a dome temperature multi-scale feature and an exhaust gas temperature multi-scale feature determined based on the operation parameter data, and perform a multi-granularity spatio-temporal attention processing on the temperature feature matrix by using a temperature prediction model to form a multi-interval context vector, and then perform a gated weight linear combination on the multi-interval context vector by using a gating mechanism, so as to accurately predict the dome temperature and the exhaust gas temperature of the hot blast stove at a future time step, and significantly improve the reliability of temperature prediction, improve the combustion efficiency and reduce the energy consumption.

[0107] In an implementation, the operation parameter data includes dome temperature time sequence data, exhaust gas temperature time sequence data, air-fuel ratio time sequence data, air flow time sequence data and gas flow time sequence data; and the feature construction module 404 is specifically configured to:

[0108] For the air-fuel ratio at any time except the first time in the air-fuel ratio time sequence data, a preset air-fuel ratio reference value is used to determine a neighboring air-fuel ratio difference value feature between the air-fuel ratio at the current time and the air-fuel ratio at the previous time, and the air-fuel ratio trend feature is constructed based on the neighboring air-fuel ratio difference value features at all times;

[0109] For any time sequence data in the dome temperature time sequence data and the exhaust gas temperature time sequence data, a window proportion corresponding to the time sequence data at different scales is determined, the mean value feature and the standard deviation feature corresponding to the time sequence data are determined according to the window proportion, and the multi-scale feature corresponding to the time sequence data is constructed, and the multi-scale feature is divided into the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature;

[0110] The dome temperature time sequence data, the exhaust gas temperature time sequence data, the air-fuel ratio time sequence data, the air flow time sequence data, the gas flow time sequence data, the dome temperature multi-scale feature and the exhaust gas temperature multi-scale feature are spliced to obtain the temperature feature matrix.

[0111] In an implementation, the temperature prediction model comprises an input layer, an improved LSTM layer, and a fully connected output layer; and the temperature prediction module 406 is specifically configured to:

[0112] receive, by the input layer, the temperature feature matrix;

[0113] extract, by the LSTM layer, the short-term context vector, the medium-term context vector, and the long-term context vector from the temperature feature matrix using the pre-trained multi-granularity spatio-temporal attention mechanism, and perform a gated weight linear combination on the short-term context vector, the medium-term context vector, and the long-term context vector to obtain a target context vector;

[0114] predict, by the fully connected output layer, the crown temperature prediction value and / or the exhaust gas temperature prediction value of the hot stove at a future time step based on the target context vector.

[0115] In an implementation, the temperature prediction module 406 is specifically configured to:

[0116] divide the hidden features of the temperature feature matrix into a short-term feature matrix, a medium-term feature matrix, and a long-term feature matrix according to a preset multi-interval ratio;

[0117] perform regional attention calculation on the short-term feature matrix, the medium-term feature matrix, and the long-term feature matrix respectively to obtain the short-term context vector, the medium-term context vector, and the long-term context vector;

[0118] determine a first weight matrix for each of the short-term context vector, the medium-term context vector, and the long-term context vector, respectively, to perform a gated weight linear combination on the short-term context vector, the medium-term context vector, and the long-term context vector to obtain a target context vector.

[0119] In an implementation, the temperature prediction module 406 is specifically configured to:

[0120] perform the following operations on any feature matrix in the short-term feature matrix, the medium-term feature matrix, and the long-term feature matrix:

[0121] perform nonlinear mapping on the hidden state of the feature matrix to obtain a mapping feature;

[0122] determine a second weight matrix for the hidden state of the mapping feature based on the mapping feature, to perform weighted aggregation on the hidden state to obtain a context vector corresponding to the feature matrix.

[0123] In an implementation, the temperature prediction module 406 is specifically configured to:

[0124] detect whether an abnormal event exists based on the short-term context vector;

[0125] If yes, the authenticity of the abnormal event is verified based on a similarity between the short-term context vector and the medium-term context vector;

[0126] In a case where the abnormal event is a real abnormal event, the short-term context vector is corrected, and a new long-term context vector is determined based on the corrected short-term context vector and the long-term context vector.

[0127] In an embodiment, the method further comprises:

[0128] If the vault temperature prediction value and / or the exhaust gas temperature prediction value is higher than the dynamic early warning threshold, temperature alarm information is generated.

[0129] The device provided by the embodiment of the present application has the same implementation principle and technical effects as the foregoing method embodiments, and for brevity of description, the part not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments.

[0130] The embodiment of the present application provides an electronic device, specifically, the electronic device comprises a processor and a storage device; the storage device stores a computer program, and the computer program performs the method of any one of the foregoing embodiments when executed by the processor.

[0131] Figure 5 The electronic device 100 provided by the embodiment of the present application comprises a processor 50, a memory 51, a bus 52 and a communication interface 53, the processor 50, the communication interface 53 and the memory 51 are connected through the bus 52; the processor 50 is used for executing the executable modules stored in the memory 51, such as a computer program.

[0132] The memory 51 can contain a high-speed random access memory (RAM), and can also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 53 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used.

[0133] The bus 52 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 5 Only one bidirectional arrow is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0134] The memory 51 is configured to store a program, and the processor 50 executes the program after receiving an execution instruction. The method executed by the device for defining a flow process according to any of the foregoing embodiments of the application can be applied to the processor 50 or implemented by the processor 50.

[0135] The processor 50 can be an integrated circuit chip with a processing capability of signals. In the implementation process, each step of the foregoing method can be completed by an integrated logic circuit of hardware in the processor 50 or an instruction in the form of software. The foregoing processor 50 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), and the like; or can be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory 51, and the processor 50 reads the information in the memory 51 and combines the hardware to complete the steps of the foregoing method.

[0136] The computer program product of the readable storage medium provided by the embodiments of the application includes a computer readable storage medium storing a program code, and the program code includes instructions for executing the method described in the foregoing method embodiments. For specific implementation, reference can be made to the foregoing method embodiments, which will not be described here.

[0137] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0138] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them. The protection scope of the present application is not limited to this. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacements to some technical features. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the temperature of the dome of a hot blast stove, characterized in that, include: The operating parameter data of the hot blast stove at the current time step are obtained. The operating parameter data includes the time series data of the dome temperature, the time series data of the exhaust gas temperature, the time series data of the air-fuel ratio, the time series data of the air flow rate, and the time series data of the gas flow rate. Based on the air-fuel ratio time series data in the operating parameter data, the air-fuel ratio trend characteristics are determined, and based on the dome temperature time series data and the exhaust gas temperature time series data in the operating parameter data, the dome temperature multi-scale characteristics and the exhaust gas temperature multi-scale characteristics are determined respectively. A temperature feature matrix is ​​constructed by combining the operating parameter data, the air-fuel ratio trend characteristics, the dome temperature multi-scale characteristics, and the exhaust gas temperature multi-scale characteristics. The temperature prediction model obtained through pre-training is used to extract the context vectors of multiple intervals corresponding to the temperature feature matrix based on the pre-trained multi-granularity spatiotemporal attention mechanism, and the context vectors of the multiple intervals are linearly combined with gating weights to predict the dome temperature and exhaust gas temperature of the hot blast stove in future time steps. The temperature prediction model includes an input layer, an improved LSTM layer, and a fully connected output layer. Using a pre-trained temperature prediction model, a pre-trained multi-granularity spatiotemporal attention mechanism is employed to extract context vectors corresponding to multiple intervals of the temperature feature matrix. These context vectors are then linearly combined with gated weights to predict the dome temperature and / or exhaust gas temperature of the hot blast stove at future time steps. This includes: The temperature feature matrix is ​​received through the input layer; By using the improved LSTM layer and a pre-trained multi-granularity spatiotemporal attention mechanism, short-term context vectors, medium-term context vectors, and long-term context vectors are extracted from the temperature feature matrix. The short-term context vector, the medium-term context vector, and the long-term context vector are then linearly combined with gated weights to obtain the target context vector. Based on the target context vector, the fully connected output layer predicts the dome temperature and / or exhaust gas temperature of the hot blast stove at future time steps.

2. The method for predicting the temperature of the hot blast stove dome according to claim 1, characterized in that; Based on the air-fuel ratio time-series data in the operating parameter data, air-fuel ratio trend characteristics are determined, and based on the dome temperature time-series data and the exhaust gas temperature time-series data in the operating parameter data, dome temperature multi-scale characteristics and exhaust gas temperature multi-scale characteristics are determined respectively. A temperature feature matrix is ​​constructed by combining the operating parameter data, the air-fuel ratio trend characteristics, the dome temperature multi-scale characteristics, and the exhaust gas temperature multi-scale characteristics, including: For the air-fuel ratio at any time other than the first time in the air-fuel ratio time series data, based on the preset air-fuel ratio reference value, the adjacent air-fuel ratio difference feature between the air-fuel ratio at the current time and the air-fuel ratio at the previous time is determined, and the air-fuel ratio trend feature is constructed based on the adjacent air-fuel ratio difference feature at all times. For any time series data in the crown temperature time series data and the exhaust gas temperature time series data, determine the window ratio corresponding to the time series data at different scales, and determine the mean feature and standard deviation feature corresponding to the time series data according to the window ratio, so as to construct the multi-scale feature corresponding to the time series data. The multi-scale feature is divided into crown temperature multi-scale feature and exhaust gas temperature multi-scale feature. The time-series data of the dome temperature, the exhaust gas temperature, the air-fuel ratio, the air flow rate, the gas flow rate, the multi-scale features of the dome temperature, and the multi-scale features of the exhaust gas temperature are concatenated to obtain a temperature feature matrix.

3. The method for predicting the temperature of the hot blast stove dome according to claim 1, characterized in that, A pre-trained multi-granularity spatiotemporal attention mechanism is used to extract short-term, medium-term, and long-term context vectors from the temperature feature matrix. A target context vector is obtained by linearly combining the short-term, medium-term, and long-term context vectors using gated weights, including: According to a preset multi-interval ratio, the hidden features of the temperature feature matrix are divided into short-term feature matrix, medium-term feature matrix and long-term feature matrix. Region attention is calculated for the short-term feature matrix, the intermediate-term feature matrix, and the long-term feature matrix respectively to obtain the short-term context vector, the intermediate-term context vector, and the long-term context vector. A first weight matrix is ​​determined for each of the short-term context vector, the intermediate context vector, and the long-term context vector, and then a gated weighted linear combination is performed on the short-term context vector, the intermediate context vector, and the long-term context vector to obtain the target context vector.

4. The method for predicting the temperature of the hot blast stove dome according to claim 3, characterized in that, Region attention is calculated for the short-term feature matrix, the intermediate-term feature matrix, and the long-term feature matrix respectively, to obtain short-term context vector, intermediate-term context vector, and long-term context vector, including: For any one of the short-term feature matrix, the medium-term feature matrix, and the long-term feature matrix, perform the following operation: The hidden states of the feature matrix are nonlinearly mapped to obtain the mapped features; Based on the mapping features, a second weight matrix is ​​determined for the corresponding hidden state, which is then used to perform weighted aggregation on the hidden state to obtain the context vector corresponding to the feature matrix.

5. The method for predicting the temperature of the hot blast stove dome according to claim 3, characterized in that, After performing region attention calculations on the short-term feature matrix, the intermediate-term feature matrix, and the long-term feature matrix respectively to obtain the short-term context vector, the intermediate-term context vector, and the long-term context vector, the method further includes: Based on the short-term context vector, detect whether there are abnormal events; If so, the authenticity of the abnormal event is verified based on the similarity between the short-term context vector and the medium-term context vector; If the abnormal event is a genuine abnormality, the short-term context vector is corrected, and a new long-term context vector is determined based on the corrected short-term context vector and the long-term context vector.

6. The method for predicting the temperature of the hot blast stove dome according to claim 1, characterized in that, The method further includes: If the predicted value of the dome temperature and / or the predicted value of the exhaust gas temperature are higher than the dynamic early warning threshold, a temperature alarm message is generated.

7. A temperature prediction device for the dome of a hot blast stove, characterized in that, include: The data acquisition module is used to acquire the operating parameter data of the hot blast stove at the current time step. The operating parameter data includes the time series data of the dome temperature, the time series data of the exhaust gas temperature, the time series data of the air-fuel ratio, the time series data of the air flow rate, and the time series data of the gas flow rate. The feature construction module is used to determine the air-fuel ratio trend feature based on the air-fuel ratio time series data in the operating parameter data, and to determine the multi-scale feature of the dome temperature and the multi-scale feature of the exhaust gas temperature based on the dome temperature time series data and the exhaust gas temperature time series data in the operating parameter data, respectively, and to construct a temperature feature matrix by combining the operating parameter data, the air-fuel ratio trend feature, the multi-scale feature of the dome temperature and the multi-scale feature of the exhaust gas temperature; The temperature prediction module is used to extract the context vectors of multiple intervals corresponding to the temperature feature matrix based on the pre-trained temperature prediction model and the pre-trained multi-granularity spatiotemporal attention mechanism, and to perform a gating weight linear combination of the context vectors of the multiple intervals to predict the dome temperature and / or exhaust gas temperature of the hot blast stove in future time steps. The temperature prediction model includes an input layer, an improved LSTM layer, and a fully connected output layer; the temperature prediction module is specifically used for: The temperature feature matrix is ​​received through the input layer; By using the improved LSTM layer and a pre-trained multi-granularity spatiotemporal attention mechanism, short-term context vectors, medium-term context vectors, and long-term context vectors are extracted from the temperature feature matrix. The short-term context vector, the medium-term context vector, and the long-term context vector are then linearly combined with gated weights to obtain the target context vector. Based on the target context vector, the fully connected output layer predicts the dome temperature and / or exhaust gas temperature of the hot blast stove at future time steps.

8. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 6.

Images