9.4.2 A case examine of CCR course of unit
It confirmed the method to discover the application of big knowledge evaluation technology by taking the continuous reforming unit CCR of a petrochemical enterprise as an example.
The refined administration of petrochemical enterprises wants increasingly collaborative administration, and collaborative administration will definitely deliver plenty of demand for correlation analysis. This demand can be between different majors inside the enterprise or between totally different enterprises. Correlation evaluation is a vital branch of massive information evaluation, which may find the correlation between variables in the chaotic knowledge. Therefore, the correlation analysis algorithm can be utilized to explore the potential elements past the normal expertise, and at last realize the potential synergies.
The specific analysis strategies can be divided into knowledge acquisition, information setting and standardization, and correlation analysis. Data acquisition is definitely the method of importing all historical information similar to operational information, quality information, corrosion data, value data, materials balance data and power data into the cloud platform (such as Aliyun). Therefore, it is essential to develop the interface between the related system and the cloud to understand real-time knowledge import. Setting and standardization of knowledge is to align the uncooked information in chronological order, such as operation data, knowledge quality, corrosion knowledge, gear operation information, price data, materials stability and vitality consumption information, in accordance with sure setting alignment algorithm. And then the info are filtered, set, rejected abnormal values, and standardized to acquire certified or legitimate data. Pearson correlation coefficient algorithm can be utilized to calculate the correlation coefficient matrix of every indicator, and to extract the variables strongly correlated with the necessary thing indicators to complete the correlation evaluation, which embody the positive correlation variables and the unfavorable correlation variables.
The Fig. 9.8 is a schematic diagram of the correlation analysis of pre-hydrogenation unit and renormalization unit including the constructive correlation variables (red) and negative correlation variables (blue) contained in. Similar strategies can be used to acquire the consequences of working conditions and uncooked material properties on product yield and tools operation, as well as different results such as operating conditions, uncooked materials properties and distillation outlet high quality on tools corrosion, production prices and environmental emissions.
Figure 9.8. (A) Results of pre-hydrotreating unit by using correlation evaluation (B) Results of CCR unit by utilizing correlation analysis.
2.Single indicator abnormal detection
Seven key indicators (such as octane quantity, energy consumption and aromatic difference index, pure hydrogen yield and thermal efficiency, flue gasoline SOX emission, sewage COD and unit cost) related to superior assessment, environmental protection and efficiency of continuous reforming unit are selected as the objects of anomaly detection. It includes a number of steps corresponding to data setting and standardization, correlation evaluation and characteristic choice, construction of prediction model and abnormal judgment of single index. Among them, characteristic choice refers again to the extraction of variables with sturdy correlation from seven key indicators after obtaining the correlation coefficient matrix.
For the establishment of prediction model, the operational variables that are strongly associated to the prediction index are chosen as the enter of the SVM prediction mannequin to establish the SVM prediction mannequin and to realize the real-time calculation of key indexes. For the irregular judgment of a single indicator, boxplot algorithm can be used to calculate the vary of every indicator and calculate the irregular restrict of every indicator. If the worth exceeds the limit, the indicator is judged to be abnormal. For example, for seven indicators of a sure enterprise’s reforming unit (unit value, octane number, effluent wastewater COD value, the SO2 emissions, power consumption and fragrant distinction index, thermal efficiency of heating furnace and pure hydrogen production rate), the enter and output of real-time information based on the SVM prediction model are shown with boxplot in the Fig. 9.9.
Figure 9.9. Box plots for dedication of outliers of seven indicators.
3.Abnormal detection of multidimensional data
In the manufacturing course of, it is possible that the entire deviates from the traditional range whereas all single indicators are normal Just like all indicators are normal in a certain human body examination, but the medical person within the state is sub-health. Therefore, it is needed to carry out abnormal detection of multi-dimension knowledge.
Abnormal detection of multi-dimension data includes information setting and standardization, extraction of characteristic variables and dimensionality reduction processing, clustering evaluation and irregular prediction and warning. Data setting and standardization are to type out the info set of seven indicates, to align them in chronological order, after which to standardize them to remove the affect of dimension and order of magnitude. In different words, the boxplot algorithm is used to gauge the abnormal points of indexes for the index knowledge. Characteristic variables and dimensionality reduction process may be extracted by principal element algorithm, in order to explain most variables with fewer variables and to realize the goal of dimensionality discount. During cluster evaluation, the principal element is extracted as the info source of clustering, and k-mens algorithm is used for clustering to seek for outliers. The method of abnormal prediction and warning is to calculate the Euclidean distance between every noticed pattern and its cluster middle. Whether the sample is anomaly or not can be judged by the gap value, whose threshold is chosen statistically based mostly on historical information.
4.Single target parameter optimization analysis
In the operation pattern library, the optimum value of the target and its corresponding strongly correlated operation variables underneath the condition of a sure sort of raw material are searched, which may mine one of the best operating experience in historical past, corresponding to mining the experience of experienced operators and solidifying it, and can be utilized complementary with optimized software corresponding to RSIM. The parameters optimization could be achieved by using raw materials clustering analysis, raw material classification mannequin, operating samples database. In the method of raw materials clustering evaluation, the historic data of the nature of the reconstituted uncooked supplies may be sorted out and reorganized. First, it goes by way of pretreatment and standardization, then principal element is used to reduce dimension, and eventually K-means clustering is adopted to output the clustering result. Based on the clustering outcomes of uncooked materials, the SVM mannequin of raw supplies classification was established, and the classification effect of the mannequin was evaluated, which may routinely classify a new batch of raw materials property data. For the optimization target strongly correlated variables, the category of uncooked material and its corresponding strongly correlated operation parameters are imported into the operation sample library, which is used as the sample of parameter optimization. In the operation sample database, the optimum values of the goal parameters under the conditions of various kinds of raw supplies are searched, as nicely as the values of the corresponding strongly correlated operation variables. Furthermore, the operation parameters could be beneficial based mostly on the raw materials quality and optimization aims.
In the operation sample database, the query statement is used to search for the value of strongly correlated operation circumstances when the target parameter is optimal beneath the condition of a certain type of raw materials. For instance, beneath the condition of e kind uncooked materials, the optimal yield of pure hydrogen is 3.392%, and the corresponding value of strongly correlated working parameters is taken.
5.Multi-objective parameter optimization analysis
According to the chosen multiple optimization goals and their optimization directions, the optimal value of each objective beneath the situation of a certain type of uncooked materials is decided, and the optimum worth and the historical actual worth are used because the coordinates of the theoretical optimum point and the actual point in the multi-dimensional area respectively. The precise level closest to the theoretical optimum is selected because the optimization result. For this objective, it’s needed to ascertain an operation pattern library first, that’s, a set of optimization variables, uncooked materials classes and their corresponding operation variables to kind a multi-objective optimization operation sample library. The clustering analysis of uncooked supplies must be accomplished to determine the corresponding uncooked material class every single day. The uncooked materials category, all optimization variables and their strongly correlated operation parameters are written into the operation pattern library in units of days. The next step is the way to determine the optimal theoretical advantage, that’s, to pick multiple optimization aims and instructions and set up the optimum value of every optimization goal beneath the situations of every kind of uncooked materials. That is to say, the optimal values of every optimization variable underneath the situation of a sure type of raw material are searched in the operation sample library, and these values are taken because the coordinates of the theoretical optimum in the multi-dimensional area. Sorting out the sample factors in numerous uncooked materials classes, that are composed of the values of variables to be optimized as coordinates, and calculating the Euclidean distance between the optimized sample factors within the multi-dimensional area beneath the condition of a certain sort of raw materials and the optimal theoretical advantages. In order to advocate operation parameters based on the scale of Euclidean distance, it is to search for the minimal value of Euclidean distance beneath the circumstances of several sorts of raw supplies and the worth of the corresponding strongly correlated operation variables within the operation pattern database, so as to realize the really helpful operation parameters based mostly on the properties of raw materials and optimization objectives. The following tables (Tables 9.2–9.4) are the optimization pattern, the optimal value of the target and the corresponding operating parameters under the condition of A class of beneficial raw materials.
Table 9.2. The Euclidean distance is sorted from small to giant, and the minimal distance pattern is decided as the optimization sample.
Sample pointFeedstock typeHydrogen yield, wt%LPG yield, wt%Fuel fuel consumption, kg/tGasoline yield, wt%Euclidean distance, m2a4. . . . . a3. . . . . a3. . . . . a3. . . . . a4. . . . . a4. . . . . a3. . . . . a4. . . . . a4. . . . . Table 9.three. The optimum worth of the target parameter.
Hydrogen yield, wt%4. LPG yield, wt%0. Fuel fuel consumption, kg/t0. Gasoline yield, wt%90. Table 9.4. Recommended working parameter.
Naphtha flow regulation of heater exchanger E701, m3/min175. Reforming response temperature, °C525. Input stress of the fourth reactor R704, MPa0. Top temperature of stabilize T701, °C58. Bottom pressure of stabilize T701, MPa0. Output temperature of furnace F701, °C527. Output temperature of furnace F704, °C535. .Unstructured knowledge analysis
Based on the text mining and evaluation of the scheduling shift log, and the correlation of structural information such as the recovery price of reformed gasoline, hydrogen production and aromatics content material of reformed gasoline, it is possible to excavate the affect of crude oil types on the recovery fee of reformed gasoline and other technical and economic indicators which may guide the crude oil procurement.
The analysis methods include textual content function analysis, transformation of unstructured data into structured information, correlation of structured information and ultimate calculation and result show. For instance, in the textual content feature evaluation stage, the historical scheduling communication class log is exported to research the text options of the crude oil species and the processing amount of the atmospheric and decompression gadget, and decide the foundations for extracting key data. Then, based on the unit of days, in accordance with the textual content characteristics decided previously, the crude oil types and corresponding processing amount are extracted and stored within the database to finish the transformation of unstructured knowledge. Continue to extract information similar to gasoline yield, hydrogen production and aromatics content material of reformed gasoline from MES and LIMS systems each day, and store them in the database after correlation with crude oil varieties to acquire relevant structured information. The weighted value of each crude oil similar to the gasoline yield and different indicators are calculated and listed from giant to small, which may information the purchase of crude oil.
7.Prediction analysis based on the properties of uncooked supplies
Based on the historic data, the prediction models of the properties of uncooked materials and the yield of gasoline, hydrogen production, dry level of gasoline, conversion fee of alkane and conversion rate of cycloalkane were established. After raw material property information being enter, the worth of the above 5 indicators may be accurately predicted to guide manufacturing.
The research strategies include data acquisition, model training, and mannequin prediction. The knowledge acquisition course of is to import all historic knowledge similar to operation data, quality knowledge, corrosion data, price data, materials balance knowledge, and vitality information into aliyun platform. The interface between the relevant system and aliyun is built to realize real-time information import. In the mannequin training stage, the raw materials high quality knowledge, and the forecast index knowledge such as gasoline yield have been derived from the uncooked data because the enter and output of SVM mannequin training. The mannequin ought to be retrained every single day to make sure the prediction accuracy. In the model prediction stage, the mannequin should predict technical and financial indicators similar to reforming gasoline yield, hydrogen manufacturing and conversion rate after 31 major laboratory evaluation information of reforming uncooked supplies are imported.
Before the enter of raw materials, the laboratory analysis knowledge of uncooked supplies are input into the prediction mannequin, and the model routinely calculates 5 technical and financial indexes of gasoline yield, hydrogen production, gasoline dry level, alkane conversion rate and cycloalkane conversion after the input of uncooked materials. Based on the forecast outcomes, technicians can adjust the relevant operating parameters in accordance with the production plan, corresponding to the blending quantity of heavy naphtha. This can save valuable sources of heavy naphtha.
The above studies have been verified on industrial devices and obtained satisfactory results. In addition, there are new findings, such because the adjustable parameters ignored by enterprises after huge information analysis. For this objective, a one-month take a look at was carried out with a special cycle of each two days, and the results were obtained.
There are normally three methods to verify the correlation between the gasoline yield and the working parameters: (1) instantaneous worth verification: the correlation between the instantaneous value of the gasoline yield and the instantaneous worth of the verification variables may be seen. This method takes under consideration the time delay issue, as a outcome of the gasoline yield can’t change immediately after the adjustment of working parameters. Therefore, transient value validation can be used as a reference for validation results. (2) MES knowledge validation: MES knowledge validation is to confirm the correlation between the gasoline yield calculated by dividing the gasoline output amount within 24 hours by the reformed feed quantity within the MES system and the mean value of the validation variables during this period. This technique eliminates the impact of time delay and is relatively accurate in calculating the gasoline yield. (3) Process simulation verification: since it is unimaginable to independently adjust a certain variable in an industrial system and maintain the properties of raw materials and other operating parameters unchanged, the method simulation method can be adopted to maintain the properties of uncooked materials and different parameters unchanged, and solely change the verification parameters to observe the change in the gasoline yield. According to the follow results, MES information validation may be selected to verify the adjustment results, and instantaneous value validation and course of simulation validation can be utilized as auxiliary references. In addition, the key parameters affecting the gasoline yield ought to be ensured to be comparatively secure so far as attainable during the validation process, in order that the validation results can actually reflect the gasoline yield affected by the validation variables.
In practice, the operating parameters won’t be adjusted greatly, and the potential content material of fragrant hydrocarbons (aromatic potential) and heavy naphtha will range greatly. Therefore, we should choose the samples for comparability by which fragrant potential and heavy naphtha refining capacity are similar.
Good follow in continuous reforming models has additionally been prolonged to other refinery items and chemical items, such as catalytic cracking and ethylene cracking. The data from different fields, such as mechanism mannequin, sample recognition, system identification, are added to the evaluation of data from multiple enterprise areas, such as the affect of course of fluctuations on the state of apparatus and so forth, in order to make the mannequin gray box.