Rectifying tower index prediction method, system and related device
By constructing time series data and setting time lag and attention weights, the delay problem in the prediction of distillation column indicators of existing models is solved, the prediction accuracy is improved and early warning is realized, and the operational stability and safety of distillation columns are enhanced.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XINJIANG ZHUNENG CHEMICAL CO LTD
- Filing Date
- 2026-01-19
- Publication Date
- 2026-04-28
AI Technical Summary
Existing empirical or linear regression models often deviate significantly from the actual results in predicting distillation column indicators, resulting in low prediction accuracy, difficulty in handling complex nonlinear relationships and long time lag effects, difficulty in capturing strong coupling between variables, and insufficient early warning lead time.
By constructing time-series data of distillation column indicators, setting time delay weights and attention weights, extracting contextual features, and utilizing a pre-set time delay perception module and attention mechanism, the distillation column indicators at future moments are predicted. This takes into account the delayed process of changes in distillation column indicators and highlights historical moments that have a greater impact on future moments.
It improves the accuracy of distillation column index prediction, reduces prediction errors caused by delays, enables early warning of potential faults and anomalies, and improves production safety and operational stability.
Smart Images
Figure CN121935541A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of distillation column technology, and in particular to a method, system and related apparatus for predicting distillation column indicators. Background Technology
[0002] An air separation unit (ASU) is an industrial device used to separate the main components of air (nitrogen, oxygen, and argon). The core component is the distillation column, whose main function is to separate liquid air into high-purity oxygen, nitrogen, and argon through cryogenic distillation.
[0003] The operational efficiency and stability of distillation columns are crucial to the quality of gas separation products. Distillation column parameters can indirectly reflect these parameters. By using empirical or linear regression models to predict these parameters, we can anticipate the operational efficiency and stability of the distillation column and allow for timely intervention. Distillation column parameters generally refer to process parameters that can be directly measured or calculated during the industrial operation and monitoring of the distillation column, such as top / bottom temperature, top / bottom pressure, and product purity.
[0004] Existing empirical models or linear regression models often have significant discrepancies between their predictions of distillation column parameters and the actual results, resulting in low prediction accuracy. Summary of the Invention
[0005] In view of the above problems, this application provides a method, system, and related apparatus for predicting distillation column indicators, so as to improve the accuracy of distillation column indicator prediction. The specific solution is as follows:
[0006] The first aspect of this application provides a method for predicting distillation column indicators, the method comprising:
[0007] Obtain distillation column index data for a period of time, and construct time series data of distillation column index based on the distillation column index data. The period of time is: a time period that includes multiple historical moments and ends with the current moment.
[0008] Time lag weights are assigned to multiple target historical moments in the time series data, wherein the multiple target historical moments correspond to target future moments.
[0009] Extract the context features of each historical moment in the time series data after setting the time delay weight, and set the attention weight of each historical moment according to the context features of each historical moment;
[0010] Predict the distillation column indices for the target future time based on time series data with the attention weights set.
[0011] In one possible implementation, the distillation column index data includes process parameters of the distillation column operation, product parameters of the distillation column, and operating status parameters of the distillation column;
[0012] The step of constructing time series data of distillation column indicators based on the distillation column indicator data includes:
[0013] The process parameters, product parameters, and operating status parameters of the distillation column at different times are obtained, and time alignment, missing value processing, and normalization are performed to obtain the time series data.
[0014] In one possible implementation, setting time-lag weights for multiple target historical moments in the time-series data includes:
[0015] Select multiple candidate historical moments based on the target future moment;
[0016] Extract the numerical and temporal features of each candidate historical moment, and calculate the degree of correlation between each candidate historical moment and the target future moment based on the numerical and temporal features of each candidate historical moment.
[0017] The multiple target historical moments are obtained by filtering from the multiple candidate historical moments based on the calculated correlation degree, and a time delay weight corresponding to the correlation degree is set.
[0018] In one possible implementation, the extraction of contextual features for each historical moment in the time series data after setting the time lag weights includes:
[0019] The forward temporal features of the time series data are extracted by traversing the time series data from the start time to the end time after setting the time delay weights.
[0020] The backward temporal features of the time series data are extracted by traversing the time series data from the end time to the start time after setting the time delay weights.
[0021] Based on the forward and backward time series features of the time series data after setting the time delay weights, the forward and backward time series features of each historical moment are concatenated to obtain the context features of each historical moment.
[0022] In one possible implementation, setting the attention weight for each historical moment based on the contextual features of each historical moment includes:
[0023] Use the contextual features of each historical moment as the key vector for each historical moment;
[0024] Generate a query vector for the time series data;
[0025] Calculate the similarity score between the key vector and the query vector at each historical moment, normalize the similarity score at each historical moment, and use the normalized similarity score at each historical moment as the attention weight at each historical moment.
[0026] One possible implementation also includes:
[0027] If the value of the distillation column index at the current moment is not within the preset range, then the changing trend of the distillation column index is determined by combining the value of the distillation column index at the current moment and the value of the distillation column index at the target future moment, and whether to trigger an early warning is determined based on the changing trend.
[0028] If the warning is triggered, the early warning time is determined. When the current time is the early warning time, the warning signal is triggered in advance. The early warning time is earlier than the target future time.
[0029] A second aspect of this application provides a distillation column index prediction system, the distillation column index prediction system comprising:
[0030] The acquisition unit is used to acquire distillation column index data over a period of time, and construct time series data of distillation column index based on the distillation column index data. The period of time is a time period that includes multiple historical moments and ends with the current moment.
[0031] A time delay unit is used to set time delay weights for multiple target historical moments in the time series data, wherein the multiple target historical moments correspond to target future moments.
[0032] The extraction unit is used to extract the context features of each historical moment in the time series data after setting the time delay weight, and set the attention weight of each historical moment according to the context features of each historical moment.
[0033] The prediction unit is used to predict the distillation column index of the target future time based on the time series data after setting the attention weight.
[0034] In one possible implementation, the distillation column index data includes process parameters of the distillation column operation, product parameters of the distillation column, and operating status parameters of the distillation column;
[0035] The acquisition unit constructs time-series data of distillation column indicators based on the distillation column indicator data, specifically configured as follows:
[0036] The process parameters, product parameters, and operating status parameters of the distillation column at different times are obtained, and time alignment, missing value processing, and normalization are performed to obtain the time series data.
[0037] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:
[0038] The memory is used to store computer programs;
[0039] The processor is used to execute the computer program so that the electronic device can implement the distillation column index prediction method of the first aspect or any implementation thereof.
[0040] The fourth aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the distillation column index prediction method of the first aspect or any implementation thereof.
[0041] Based on the above technical solution, this application provides a method, system, and related apparatus for predicting distillation column indicators. When predicting distillation column indicators for future times from the current time, this method obtains distillation column indicator data for a period of time, including multiple historical timestamps from the current time, constructing time series data of distillation column indicators. Time lag weights are assigned to target historical times in the time series data that correspond to the target future time. This highlights the target historical times in the time series data that influence changes in distillation column indicators for future times. Then, the contextual features of each historical timetamp in the time series data with assigned time lag weights are extracted. Attention weights are assigned to each historical timetamp based on its contextual features, highlighting historical times important for predicting distillation column indicators for future times. Finally, the distillation column indicators for future times are predicted based on the time series data with assigned attention weights. This method considers the delayed process of distillation column index changes when predicting distillation column indexes. Instead of selecting the current moment, it selects the historical moment that truly affects the future changes in distillation column indexes. By using time delay weights and attention weights, it highlights the historical moments that have a greater impact on the future changes in distillation column indexes, effectively reducing the prediction error caused by the delay in distillation column index changes and improving the prediction accuracy of distillation column indexes. Attached Figure Description
[0042] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.
[0043] Figure 1A flowchart illustrating a method for predicting distillation column indicators provided in this application embodiment;
[0044] Figure 2 A schematic diagram of a distillation column index prediction system provided in this application embodiment;
[0045] Figure 3 This application provides a hardware structure block diagram of an electronic device. Detailed Implementation
[0046] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.
[0047] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.
[0048] The terms "first," "second," etc., used in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.
[0049] In a distillation column, a local change does not immediately cause a change in the column's parameters; rather, there is a "delay" process. For example, a change in the reflux ratio of a distillation column requires several gas-liquid equilibrium stages for the composition of the gas at the top of the column to change significantly, and the purity may take several minutes or even tens of minutes to improve. This "delay" is essentially determined by the material residence time, heat transfer rate, and the dynamic process of gas-liquid interaction.
[0050] Empirical models or linear regression models, when making predictions, ignore this lag process and choose to predict directly based on local changes at the current moment. This leads to a discrepancy between the predicted and actual values, with predictions either leading or lagging behind. For example, if the feed flow rate increases at 10:00, causing the product concentration to begin decreasing at 10:10, the empirical or linear regression model, predicting the change in product concentration directly based on the current feed flow rate, will have an error in its prediction because the feed flow rate varies at different times, resulting in a discrepancy between the predicted result and the actual decrease in product concentration at 10:10.
[0051] Therefore, when predicting distillation column indicators using empirical models or linear regression models, the prediction accuracy is low, they cannot handle complex nonlinear relationships and long time lag effects, they are difficult to capture strong coupling between variables, and the early warning lead time is insufficient, making it difficult to respond to faults or anomalies in a timely manner.
[0052] To address the aforementioned problems, this application provides a method for predicting distillation column performance indicators. The method for predicting distillation column performance indicators according to this application will be described in detail below with reference to the accompanying drawings.
[0053] Reference Figure 1 , Figure 1 This is a flowchart illustrating a method for predicting distillation column indicators provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiment of this application, a method for predicting distillation column indicators may include steps S10 to S13, which are described in detail below.
[0054] S10. Obtain distillation column index data for a period of time, and construct time series data of distillation column index based on the distillation column index data. A period of time is a time period that includes multiple historical moments and ends with the current moment.
[0055] In distillation column parameters, the parameters refer to the process parameters that can be directly measured or calculated during the industrial operation and monitoring of the distillation column. These parameters can be categorized into two types: product parameters and operating parameters. Product parameters may include: top / bottom temperature, top / bottom pressure, and product purity (concentration of products such as oxygen, nitrogen, and argon). Operating parameters may include: reflux ratio, feed flow rate, and reboiler load. Changes in operating parameters can lead to changes in product parameters.
[0056] Therefore, in this embodiment, the distillation column index data can refer to the parameter values of various distillation column indicators, which are reclassified into three categories: process parameters of distillation column operation, product parameters of distillation column, and operating status parameters of distillation column. Specifically, in this embodiment, sensors in the distillation column can be used to collect key distillation column indicators during operation, such as top / bottom temperature, top / bottom pressure, reflux ratio, feed flow rate, and reboiler load. The sampling method is second-level sampling, with sampling intervals typically one or several seconds, to obtain the process parameters of the distillation column operation in this embodiment. The composition parameters of the distillation column products, such as the concentrations of oxygen, nitrogen, and argon, are collected using an online analyzer, with sampling intervals typically 1-10 minutes, to obtain the product parameters of the distillation column in this embodiment. By recording data, parameters such as the start-up and shutdown status, operating model, operating conditions (external weather conditions, etc.), and production safety information of the distillation column are obtained, thus obtaining the operating status parameters of the distillation column in this embodiment.
[0057] After collecting three types of distillation column index data—process parameters, product parameters, and operating status parameters—at different times, the distillation column index data needs to be preprocessed, which may include time alignment, missing value handling, and normalization.
[0058] For time alignment, since the parameters have different sampling frequencies, it is necessary to perform time alignment on the multi-source distillation column index data with different sampling frequencies. Fast-sampled (high sampling frequency, relatively short sampling interval) distillation column index data is downsampled to the same time step (e.g., 10 seconds or 30 seconds), and slow-sampled (low sampling frequency, relatively long sampling interval) distillation column index data is upsampled to the same time step. Specifically, this embodiment can use specific rules to filter or merge fast-sampled distillation column index data to obtain data with the same time interval as slow-sampled distillation column index data. Alternatively, it can generate new slow-sampled distillation column index data using completion rules to obtain data with the same time interval as fast-sampled distillation column index data. Alternatively, it can process both fast-sampled and slow-sampled distillation column index data simultaneously to achieve the same median time interval.
[0059] For handling missing values, this embodiment can use methods such as forward padding, linear interpolation, or Kalman filtering to fill in the missing data and ensure data integrity.
[0060] Regarding normalization, this embodiment can unify the scale of the distillation column index data by changing the scale of the data, thus avoiding the impact of scale differences between data on the prediction effect. Specifically, this embodiment can use min-max normalization (linearly mapping the data to a specified interval using a formula) to scale the various parameters in the distillation column index data to a uniform interval, such as the 0-1 interval or the -1 to 1 interval.
[0061] After the above data preprocessing, time series data of distillation column index data (data items arranged in chronological order) can be obtained.
[0062] S11. Set time lag weights for multiple target historical moments in the time series data, where multiple target historical moments correspond to target future moments.
[0063] In this embodiment, a preset time-lag perception module can be used to set time-lag weights for multiple target historical moments in the time series data. The preset time-lag perception module can refer to a module that selects the lag features most relevant to the prediction time using a soft weighting mechanism. In this embodiment, the target future moment can refer to the future moment to be predicted, and the target historical moment can refer to historical moments that influence or contribute to the distillation column indicators of the predicted future moment. Since the changes in distillation column indicators have a delay, this embodiment can use the preset time-lag perception module to determine which historical moments are more important to the distillation column indicators of the predicted future moment from multiple historical moments at the current moment. The time-lag weight can refer to the degree of importance of the historical moments reflecting the time lag in predicting the distillation column indicators of the future moment.
[0064] Specifically, the process of setting the time delay weight in the preset time delay perception module can be as follows:
[0065] The preset time delay perception module selects multiple candidate historical moments based on the target future moment; extracts the numerical and temporal features of each candidate historical moment, and calculates the correlation between each candidate historical moment and the target future moment based on the numerical and temporal features of each candidate historical moment; selects multiple target historical moments from the multiple candidate historical moments based on the calculated correlation, and sets time delay weights corresponding to the correlation.
[0066] Here, candidate historical moments refer to historical moments that may affect the distillation column indicators at the target future moment. The preset time-delay perception module, based on the delay characteristics of changes in the distillation column indicators at the target future moment, constructs multiple lag channels to determine multiple candidate historical moments for the target future moment. For example, if the target future moment is ten minutes after the current moment, two lag channels can be constructed: one for 5 minutes ago and one for 10 minutes ago, using the historical moments 5 minutes and 10 minutes ago as candidate historical moments. Similarly, if the target future moment is one hour after the current moment, two lag channels can be constructed: one for 30 minutes ago and one for 40 minutes ago, using the historical moments 30 minutes and 40 minutes ago as candidate historical moments. Furthermore, since the delay in changes to operating parameters differs from the delay in changes to product parameters, the preset time delay perception module can use the data type of the processed data as auxiliary data for constructing the lag channel. For example, when constructing a lag channel for operating parameters in distillation column index data, the key lag time is within the time range of 5-20 minutes, while when constructing a lag channel for product parameters in distillation column index data, the key lag time is within the time range of 30-60 minutes.
[0067] Of course, the preset time delay perception module can also determine multiple predictable target future moments based on the time window of the time series data (the time span of multiple historical moments). For example, if the time window of the time series data is 1 hour, it means that the time series data can include all distillation column index data in the hour before the current moment. Based on the delay time of the changes in the distillation column indexes, it can predict the distillation column indexes in short-term time periods such as 1 minute, 2 minutes, and 5 minutes after the current moment, as well as in medium-term time periods such as 10 minutes and 20 minutes after the current moment, and in long-term time periods such as 30 minutes and 60 minutes after the current moment.
[0068] The soft weighting mechanism of the preset time delay perception module can refer to assigning continuously changing weights to the processing objects based on their data characteristics, similarity, and other indicators, so as to flexibly adjust the importance or contribution of the processing objects in the application scenario. Therefore, in this embodiment, the preset time delay perception module can extract the numerical and temporal features of each candidate historical moment, thereby dynamically determining the time delay weight of each candidate historical moment. The numerical features of the candidate historical moment can refer to the characteristics of the parameter values of the candidate historical moment, such as the parameter values and differences. In this embodiment, the difference can be calculated using local data from the candidate historical moment and adjacent moments before and after it, and used as the difference for that candidate historical moment. The temporal features of the candidate historical moment can refer to its characteristics in the time dimension, such as the time interval with the target future moment.
[0069] The correlation between a candidate historical moment and a target future moment can refer to the potential influence or contribution of the candidate historical moment to the target future moment. In this embodiment, the preset time-delay perception module can construct a feature vector for each candidate historical moment from its numerical and temporal features. These feature vectors are then combined into a feature matrix. The preset time-delay perception module performs a linear transformation with this feature matrix using a built-in learnable weight matrix to obtain a correlation score for each candidate historical moment, which serves as the correlation degree for that moment. Of course, after obtaining the correlation score for each candidate historical moment, further nonlinear mapping can be applied using activation functions to strengthen the effective correlation.
[0070] After obtaining the correlation degree of each candidate historical moment, the pre-set time-delay perception module normalizes the correlation degree of all candidate historical moments and selects the target historical moment from the candidate historical moments according to the correlation degree. Corresponding time-delay weights are assigned to the target historical moment and candidate historical moments. Historical moments with high correlation (target historical moments) can be assigned higher time-delay weights, and historical moments with low correlation (other candidate historical moments) can be assigned lower time-delay weights. The specific assignment rules can be dynamically learned and determined by the pre-set time-delay perception module. Furthermore, after determining the target historical moment, the pre-set time-delay perception module can compare the mean square error of the time period containing the target historical moment with that of a random time period through occlusion experiments to prove that the target historical moment is indeed crucial for future prediction.
[0071] This embodiment captures the time lag effect of distillation column indicators through a preset time lag sensing module, and highlights historical moments related to the time lag of distillation column indicators at the target future time by using time lag weighting, thereby reducing prediction bias caused by delay.
[0072] S12. Extract the context features of each historical moment in the time series data after setting the time lag weight, and set the attention weight of each historical moment according to the context features of each historical moment.
[0073] S13. Predict the distillation column index of the target at future moments based on the time series data after setting attention weights.
[0074] Among them, contextual features can refer to auxiliary features such as time features and local sequence features that describe historical moments, and a set of features that reflect the potential relationship between historical moments and other moments.
[0075] When time-lag weights are set in time-series data, these weights can guide the effective extraction of contextual features from historical moments, focusing more on historical moments with high time-lag weights and highlighting their information (such as their time-series position), while reducing the interference from historical moments with low time-lag weights. Specifically, the process of extracting contextual features from each historical moment in time-series data after setting time-lag weights can be as follows:
[0076] The time series data is traversed from start to end time according to the time sequence after setting time delay weights to extract forward time series features; the time series data is then traversed from end to start time according to the time sequence after setting time delay weights to extract backward time series features; based on the forward and backward time series features of the time series data after setting time delay weights, the forward and backward time series features of each historical moment are concatenated to obtain the context features of each historical moment.
[0077] In this embodiment, bidirectional feature extraction of time series data after setting time delay weights can obtain more stable global features of time series data and effectively avoid deviations at a single time step.
[0078] After extracting the bidirectional features of the time series data with time lag weights, it becomes easier for the attention mechanism to dynamically weight the feature importance at different historical moments. The process of setting the attention weight for each historical moment based on its contextual features can be described as follows:
[0079] The contextual features of each historical moment are used as the key vector of each historical moment; a query vector for the time series data is generated; the similarity score between the key vector and the query vector of each historical moment is calculated, and the similarity score of each historical moment is normalized. The normalized similarity score of each historical moment is used as the attention weight of each historical moment.
[0080] The query vector can be based on the context features of the current moment, or it can be a global aggregation of the context features of all historical moments, with the aggregation result used as the query vector. This embodiment focuses on the context features of historical moments with high attention weights for time-series data with attention weights set, while weakening the context features of historical moments with low attention weights, thereby achieving prediction of distillation column indicators for the target future moment.
[0081] This embodiment, through final attention weighting, can automatically identify and focus on the most important historical moments (within the entire historical data window, state changes at historical moments contribute more to the prediction results) when predicting distillation column indicators for target future moments, effectively improving prediction accuracy. Furthermore, this embodiment can simultaneously predict distillation column indicators for multiple target future moments (based on data from corresponding target historical moments), and in addition to predicting the distillation column indicators for target future moments, it can also obtain the probability of exceeding limits for the corresponding target future moments, used to provide early warning of potential anomalies. By predicting distillation column indicators in advance, this embodiment can identify potential faults or safety hazards in advance, effectively reducing manual monitoring costs and improving production safety.
[0082] After obtaining the predicted distillation column index for the target future time, this embodiment can determine whether an early warning is needed by combining the current distillation column index and the predicted target future distillation column index. Specifically, if the current distillation column index value is not within the preset range, the changing trend of the distillation column index is determined by combining the current value and the target future value, and an early warning is triggered based on the changing trend. If an early warning is triggered, an advance warning time is determined, which is earlier than the target future time. When the current time is the advance warning time, the early warning signal is triggered in advance. Furthermore, this embodiment can set different warning levels based on the difference between the target future distillation column index and the maximum or minimum value of the preset range, such as yellow, orange, and red warnings. Different warning levels can correspond to different warning response measures. Through precise early warning and optimized control, this embodiment can effectively improve the stability of distillation column operation and reduce downtime and energy waste caused by abnormalities.
[0083] The preset range refers to the reasonable range of values for distillation column indicators, which can be determined based on the historical data distribution of these indicators. If the current value of the distillation column indicator is outside the preset range, it indicates an anomaly. By combining this with the predicted future value of the indicator, the trend of change can be determined, thus deciding whether to set an alert. Specifically, if the current value of the distillation column indicator exceeds the maximum value of the preset range, but the value of the indicator in the target future time is within the preset range, it indicates a short-term increase in the indicator, but its trend is downward, eventually returning to the normal range. Conversely, if the current value of the distillation column indicator is less than the minimum value of the preset range, but the value of the indicator in the target future time is within the preset range, it indicates a short-term decrease in the indicator, but its trend is upward, eventually returning to the normal range. In both of these cases, no alert is required. If the current value of the distillation column indicator exceeds the maximum value of the preset range, and the value of the distillation column indicator at the target future time is also greater than the maximum value of the preset range, or if the current value of the distillation column indicator is less than the minimum value of the preset range, and the value of the distillation column indicator at the target future time is also less than the minimum value of the preset range, both of these situations indicate that the distillation column indicator will remain in an abnormal state after a period of time, and an early warning is required. Alternatively, in another optional embodiment, if the value of the distillation column indicator at the target future time does not fall within the preset range, an early warning signal can be triggered directly.
[0084] When this embodiment determines that an early warning is needed based on the changing trend of the distillation column indicators, the specific method for determining the early warning time can be as follows: compare the time difference between the actual time of exceeding the limit and the target future time, determine the early warning lead time (which can be less than the time difference), and determine the early warning time based on the early warning lead time. The actual time of exceeding the limit can refer to the moment when the value of the distillation column indicator first fails to fall within the preset range during operation; this moment originates from real-time data collected online.
[0085] Furthermore, this embodiment can achieve the prediction process of the above-mentioned distillation column indicators by training a neural network model.
[0086] The structure of a neural network model can include a time-lag perception module, a BiLSTM encoder, and a self-attention module. The time-lag perception module constructs multiple lag channels for each input feature and selects the lag features most relevant to the training objective through a soft weighting mechanism. This allows the model to dynamically learn which time lags are more important for distillation column prediction. BiLSTM (Bidirectional Long Short-Term Memory) is a recurrent neural network used to capture long-term dependencies in time series. It considers both past and future information, enabling the model to more comprehensively understand the relationships between various variables (distillation column indicators) during the distillation process. The self-attention module uses dynamic weighting to reflect the importance of features at different time steps, helping the model effectively extract features from strongly coupled variables and long-term dependencies.
[0087] The process of constructing the training dataset can be described as follows:
[0088] For data acquisition: Process data of the distillation column is collected: key variables of the distillation column are collected through sensors, including: top / bottom temperature, top / bottom pressure, feed flow rate, reflux ratio, reboiler load, etc. All of the above data are sampled at the second level. Online analyzer data acquisition: the composition of the distillation column product is collected, with sampling intervals of 1-10 minutes. Status and operating condition data of the distillation column are collected: the start-up and shutdown status, operating mode, operating conditions (such as external weather conditions), and production safety information of the unit in the distillation column are recorded.
[0089] For data preprocessing, time step alignment and missing value processing are first performed on multi-source data with different sampling frequencies to obtain raw data. Then, key features are extracted and constructed from the raw data, such as moving averages, standard deviations, slopes, and liquid level / temperature differences. Lag features are introduced as additional features, such as data from the past 30 or 60 minutes, to enhance the model's ability to capture historical dependencies. Specifically, for the process data of the distillation column in the raw data (using high-frequency sampling values at the second or sub-minute level), moving averages, standard deviations, and slopes are calculated for data within a set time window (e.g., 1 minute, 5 minutes, or 10 minutes) to characterize the rapid fluctuation trend and short-term dynamic characteristics of the distillation column process data. For the online analyzer data (using low-frequency data sampled from 1 to 10 minutes), moving averages, standard deviations, and slopes are calculated for data within a longer time window (e.g., 30 minutes or 60 minutes) to characterize the fluctuation trend and steady-state characteristics of the online analyzer data. Finally, the calculated numerical data and the original data are normalized to obtain the training dataset of the model. The dataset is divided into training set (60%-70%), validation set (15%-20%) and test set (15%-20%) according to the time order of the data to ensure that there is no data omission among the training set, validation set and test set.
[0090] After obtaining the training dataset, the time-series data in the training set is input into the model for training. The time-lag-aware module performs time-lag-aware weighting before inputting it into the BiLSTM encoder. The time-series features output by the BiLSTM encoder are weighted through a self-attention mechanism, dynamically weighting different time steps, allowing the model to automatically identify and focus on the most important time steps for prediction. During model training, mean squared error can be used as the loss function for regression problems, supplemented by binary classification loss to predict whether an alert is triggered. The model can be trained using the AdamW optimizer (an efficient optimizer), with OneCycle learning rate (a dynamic learning rate scheduling strategy) used for dynamic adjustment of the learning rate to improve training performance. This embodiment also incorporates a Dropout layer (a technique to prevent overfitting and improve generalization ability) to reduce overfitting, and uses L2 regularization (a parameter regularization technique) to avoid excessively large parameters. By monitoring training and validation errors, an early stopping strategy is used to prevent overfitting, and the optimal model is saved through model checkpoints.
[0091] The time-delay awareness module can construct multiple time-delay channels and learn the learnable weights of each channel, aligning these weights by time for easier integration with subsequent attention. During the learning process, the time-delay awareness module can disable certain delayed channels and observe whether the model's predictive performance degrades, thereby quantifying the contribution of each delayed channel to the model's prediction. The self-attention module can indirectly update the attention weights through backpropagation of the prediction error, allowing historical moments that contribute more to the prediction result to receive higher attention weights. Furthermore, to improve interpretability and stability, this embodiment introduces entropy constraints and sparsity regularization for the attention distribution during training, making the model's attention distribution more focused and readable.
[0092] The trained model can be deployed to the real-time monitoring system of the distillation column. Data acquisition and processing are performed using a PLC (Programmable Logic Controller) or DCS (Distributed Control System). Real-time data is transmitted to the trained model via an API (Application Programming Interface) for processing, enabling rolling predictions of multiple distillation column indicators. Specifically, the system can input the latest historical data window (e.g., data from one hour prior to the current moment) into the model to generate predictions for multiple future distillation column indicators. Furthermore, the system continuously monitors the model's prediction accuracy and warning accuracy during operation. If model performance deteriorates or operating conditions change, the model can be optimized through periodic incremental training or online learning. The system automatically records the results of each prediction and warning, including prediction error, warning response time, and threshold adjustments, for subsequent analysis and optimization by operators. The model is also periodically retrained based on operational experience and new data, allowing it to adapt to changes in the distillation column and production conditions.
[0093] This application provides a method for predicting distillation column indicators. When predicting distillation column indicators for future times from the current time, this method obtains distillation column indicator data for a period of time, including multiple historical timestamps from the current time, to construct time series data of distillation column indicators. A time lag weight is assigned to the target historical timestamps corresponding to the target future time. This highlights the target historical timestamps in the time series data that have an impact on the changes in distillation column indicators for future times. The contextual features of each historical timetamp in the time series data with the assigned time lag weights are then extracted. An attention weight is assigned to each historical timetamp based on its contextual features, highlighting the historical timestamps that are important for predicting the distillation column indicators for future times. The distillation column indicators for future times are then predicted based on the time series data with the assigned attention weights. This method considers the delayed process of distillation column indicator changes during indicator prediction. Instead of selecting the current timetamp, it selects the historical timestamps that truly influence the changes in distillation column indicators for future times. By using time lag weights and attention weights to highlight the historical timestamps that have a greater impact on the changes in distillation column indicators for future times, it effectively reduces prediction errors caused by the delay in changes in distillation column indicators, thus improving the prediction accuracy of distillation column indicators.
[0094] The above describes a method for predicting distillation column indicators provided by embodiments of this application. The following describes the apparatus for performing the above method for predicting distillation column indicators.
[0095] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a distillation column index prediction system provided in an embodiment of this application, as shown below. Figure 2 As shown, the distillation column index prediction system may include:
[0096] The acquisition unit 100 is used to acquire distillation column index data over a period of time and construct time series data of distillation column index based on the distillation column index data. A period of time is a time period that includes multiple historical moments and ends with the current moment.
[0097] The time delay unit 110 is used to set time delay weights for multiple target historical moments in time series data, wherein the multiple target historical moments correspond to target future moments.
[0098] Extraction unit 120 is used to extract the context features of each historical moment in the time series data after setting time lag weights, and set the attention weight of each historical moment according to the context features of each historical moment.
[0099] Prediction unit 130 is used to predict the distillation column index of the target at future times based on time series data with attention weights set.
[0100] In one possible implementation, the distillation column index data includes process parameters of the distillation column operation, product parameters of the distillation column, and operating status parameters of the distillation column.
[0101] The time series data of distillation column indicators constructed from the distillation column indicator data in unit 100 can be specifically configured as follows:
[0102] The process parameters, product parameters, and operating status parameters of the distillation column at different times are obtained, and time alignment, missing value handling, and normalization are performed to obtain time series data.
[0103] In one possible implementation, the process of setting time delay weights for multiple target historical moments in the time series data in the time delay unit 110 can be specifically configured as follows:
[0104] Select multiple candidate historical moments based on the target future moment; extract the numerical and temporal features of each candidate historical moment, and calculate the correlation between each candidate historical moment and the target future moment based on the numerical and temporal features of each candidate historical moment; filter out multiple target historical moments from the multiple candidate historical moments based on the calculated correlation, and set time delay weights corresponding to the correlation.
[0105] In one possible implementation, the extraction unit 120 extracts the context features of each historical moment in the time series data after setting time delay weights, which can be specifically configured as follows:
[0106] The time series data is traversed from start to end time according to the time sequence after setting time delay weights to extract forward time series features; the time series data is then traversed from end to start time according to the time sequence after setting time delay weights to extract backward time series features; based on the forward and backward time series features of the time series data after setting time delay weights, the forward and backward time series features of each historical moment are concatenated to obtain the context features of each historical moment.
[0107] In one possible implementation, the extraction unit 120 sets the attention weight for each historical moment based on the contextual features of each historical moment, which can be specifically configured as follows:
[0108] The contextual features of each historical moment are used as the key vector of each historical moment; a query vector for the time series data is generated; the similarity score between the key vector and the query vector of each historical moment is calculated, and the similarity score of each historical moment is normalized. The normalized similarity score of each historical moment is used as the attention weight of each historical moment.
[0109] In one possible implementation, the distillation column index prediction system may also include an early warning unit:
[0110] The early warning unit is used to determine the trend of the distillation column index if the value of the current distillation column index is not within the preset range. It combines the value of the current distillation column index with the value of the distillation column index at the target future time, and determines whether to trigger an early warning based on the trend. If an early warning is triggered, it determines the early warning time. When the current time is the early warning time, the early warning signal is triggered in advance. The early warning time is earlier than the target future time.
[0111] This application also provides an electronic device in its embodiments. (See reference...) Figure 3 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.
[0112] like Figure 3 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. When the electronic device is powered on, the RAM 303 also stores various programs and data required for the operation of the electronic device. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output interface (I / O interface) 305 is also connected to the bus 304.
[0113] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, memory cards, hard drives, etc.; and communication devices 309. Communication device 309 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.
[0114] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the distillation column index prediction methods provided in this application.
[0115] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the distillation column index prediction methods provided in this application.
[0116] It should also be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the system embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.
[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0118] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.
[0119] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).
[0120] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0121] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.
[0122] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for predicting distillation column indicators, characterized in that, The method for predicting distillation column indicators includes: Obtain distillation column index data for a period of time, and construct time series data of distillation column index based on the distillation column index data. The period of time is: a time period that includes multiple historical moments and ends with the current moment. Time lag weights are assigned to multiple target historical moments in the time series data, wherein the multiple target historical moments correspond to target future moments. Extract the context features of each historical moment in the time series data after setting the time delay weight, and set the attention weight of each historical moment according to the context features of each historical moment; Predict the distillation column indices for the target future time based on time series data with the attention weights set.
2. The method for predicting distillation column indicators according to claim 1, characterized in that, The distillation column index data includes process parameters of the distillation column operation, product parameters of the distillation column, and operating status parameters of the distillation column. The step of constructing time series data of distillation column indicators based on the distillation column indicator data includes: The process parameters, product parameters, and operating status parameters of the distillation column at different times are obtained, and time alignment, missing value processing, and normalization are performed to obtain the time series data.
3. The method for predicting distillation column indicators according to claim 1, characterized in that, The step of setting time lag weights for multiple target historical moments in the time series data includes: Select multiple candidate historical moments based on the target future moment; Extract the numerical and temporal features of each candidate historical moment, and calculate the degree of correlation between each candidate historical moment and the target future moment based on the numerical and temporal features of each candidate historical moment. The multiple target historical moments are obtained by filtering from the multiple candidate historical moments based on the calculated correlation degree, and a time delay weight corresponding to the correlation degree is set.
4. The method for predicting distillation column indicators according to claim 1, characterized in that, The extraction of contextual features for each historical moment in the time series data after setting the time lag weights includes: The forward temporal features of the time series data are extracted by traversing the time series data from the start time to the end time after setting the time delay weights. The backward temporal features of the time series data are extracted by traversing the time series data from the end time to the start time after setting the time delay weights. Based on the forward and backward time series features of the time series data after setting the time delay weights, the forward and backward time series features of each historical moment are concatenated to obtain the context features of each historical moment.
5. The method for predicting distillation column indicators according to claim 1, characterized in that, The step of setting the attention weight for each historical moment based on the contextual features of each historical moment includes: Use the contextual features of each historical moment as the key vector for each historical moment; Generate a query vector for the time series data; Calculate the similarity score between the key vector and the query vector at each historical moment, normalize the similarity score at each historical moment, and use the normalized similarity score at each historical moment as the attention weight at each historical moment.
6. The method for predicting distillation column indicators according to claim 1, characterized in that, Also includes: If the value of the distillation column index at the current moment is not within the preset range, then the changing trend of the distillation column index is determined by combining the value of the distillation column index at the current moment and the value of the distillation column index at the target future moment, and whether to trigger an early warning is determined based on the changing trend. If the warning is triggered, the early warning time is determined. When the current time is the early warning time, the warning signal is triggered in advance. The early warning time is earlier than the target future time.
7. A distillation column index prediction system, characterized in that, The distillation column index prediction system includes: The acquisition unit is used to acquire distillation column index data over a period of time, and construct time series data of distillation column indexes based on the distillation column index data. The period of time is a time period that includes multiple historical moments and ends with the current moment. A time delay unit is used to set time delay weights for multiple target historical moments in the time series data, wherein the multiple target historical moments correspond to target future moments. The extraction unit is used to extract the context features of each historical moment in the time series data after setting the time delay weight, and to set the attention weight of each historical moment according to the context features of each historical moment. The prediction unit is used to predict the distillation column index of the target future time based on the time series data after setting the attention weight.
8. The distillation column index prediction system according to claim 7, characterized in that, The distillation column index data includes process parameters of the distillation column operation, product parameters of the distillation column, and operating status parameters of the distillation column. The acquisition unit constructs time-series data of distillation column indicators based on the distillation column indicator data, specifically configured as follows: The process parameters, product parameters, and operating status parameters of the distillation column at different times are obtained, and time alignment, missing value processing, and normalization are performed to obtain the time series data.
9. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the distillation column index prediction method as described in any one of claims 1 to 6.
10. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the distillation column index prediction method as described in any one of claims 1 to 6.