Pyrolysis gasoline hydrogenation yield mutation prediction method considering time sequence characteristics

Abstract

The invention discloses a pyrolysis gasoline hydrogenation yield mutation prediction method considering time sequence characteristics, and the method comprises the following steps: S1, collecting pyrolysis gasoline hydrogenation real-time data; S2, inputting the real-time data into a pyrolysis gasoline hydrogenation yield mutation prediction model which is constructed by the following steps: S3, outputting a yield prediction result by the model. Prediction is carried out based on the method, the fitting capacity of mutation points is enhanced, the prediction error is small when the yield is mutated, and the practicability of the model is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent production, and specifically relates to a method for predicting the sudden change of the hydrogenation yield of pyrolysis gasoline considering temporal characteristics. Background Art

[0002] Pyrolysis gasoline is an important intermediate product in petrochemical industry, and its main components include olefins, aromatics, naphthenes, paraffins, etc.

[0003] The hydrogenation process converts unsaturated hydrocarbons (such as olefins) in pyrolysis gasoline into saturated hydrocarbons through catalytic hydrogenation reaction to improve the product quality. Accurate yield prediction is the key to optimizing the hydrogenation process, improving resource utilization efficiency, product quality and economic benefits.

[0004] In the prior art, to predict the yield, the XGBoost algorithm is adopted: the main components of pyrolysis gasoline (olefin content, aromatic content, naphthene content, paraffin content, other impurity content) are used as features; the XGBoost model is trained with historical production data, and the target variable is the hydrogenation yield.

[0005] XGBoost is an efficient gradient boosting tree algorithm, which is widely used in prediction and classification problems. Advantages: good at dealing with non-linear relationships, strong ability, supporting feature importance evaluation, suitable for medium and small-scale data sets; Disadvantages: for particularly large data sets, it may require a large amount of memory and computing time; sensitive to outliers, such as extreme outliers may affect the performance of the model.

[0006] Specifically in yield prediction, the XGBoost algorithm has the following defects:

[0007] (1) Insufficient fitting ability for mutation points

[0008] When the yield suddenly changes, the prediction error is large, which affects the practicality of the model.

[0009] (2) Insufficient feature engineering

[0010] The existing features (such as olefin content, aromatic content, naphthene content, paraffin content, other impurity content) may not be able to fully describe the reasons for the sudden change of yield. There is a lack of in-depth analysis of the dynamic changes of operating conditions.

[0011] (3) Unbalanced data distribution

[0012] The data of mutation points accounts for a small proportion in the training set, and the model tends to fit the mainstream data distribution. As a result, the prediction ability of the model for mutation points is insufficient. Summary of the Invention

[0013] The present invention addresses the problems existing in the background art and proposes a method for predicting the sudden change of the hydrogenation yield of pyrolysis gasoline considering temporal characteristics.

[0014] Technical solution:

[0015] A method for predicting the sudden change of the hydrogenation yield of pyrolysis gasoline considering temporal characteristics, which comprises the following steps:

[0016] S1. Collect real-time data of pyrolysis gasoline hydrogenation:

[0017] S2. Input the real-time data into the prediction model for the sudden change of the hydrogenation yield of pyrolysis gasoline, and the model is constructed through the following steps:

[0018] S2-1. Obtain historical data of pyrolysis gasoline hydrogenation;

[0019] S2-2. Data preprocessing;

[0020] S2-3. Obtain all data samples based on the preprocessed data, including: ① Using the SMOTE algorithm to oversample the minority class samples; ② Merging with the normal data samples to generate all data samples; ③ Constructing a feature library;

[0021] S2-4. Divide all data samples into a training set and a test set;

[0022] S2-5. Construct a long short-term memory network (LSTM) model, train and test it to obtain a prediction model for the sudden change of the hydrogenation yield of pyrolysis gasoline;

[0023] S3. The model outputs the yield prediction result.

[0024] Preferably, the model is retrained according to the historical data of the cycle.

[0025] Preferably, the data preprocessing includes:

[0026] (1) Standardization or normalization;

[0027] (2) Time series segmentation, converting the data into a three-dimensional array (number of samples, time step, number of features) suitable for LSTM input;

[0028] (3) Data cleaning.

[0029] Preferably, the SMOTE algorithm is used to oversample the minority class samples, which specifically includes the following steps:

[0030] (1) Set SMOTE parameters;

[0031] (2) Perform oversampling, specifically including:

[0032] 1) For each minority class sample, find its k nearest neighbor samples;

[0033] 2) Randomly select a nearest neighbor sample and generate a new synthetic sample between the two;

[0034] 3) Repeat the above process until the required number of samples is reached;

[0035] (3) Conduct a quality inspection on the synthetic samples.

[0036] The feature library includes: main component features, operation features, catalyst features, environmental features, composite features, and time series features. Specifically:

[0037] The main component features include: olefin content, aromatic hydrocarbon content, naphthene content, paraffin content, and impurity content;

[0038] The operation features include: reaction temperature, reaction pressure, reaction time, and hydrogen-oil ratio;

[0039] The catalyst features include: catalyst activity index, catalyst usage duration, and catalyst regeneration times;

[0040] The environmental features include: environmental temperature and atmospheric humidity;

[0041] The composite features include: temperature-pressure interaction effect, hydrogen-oil ratio adjustment factor, and catalytic efficiency index;

[0042] The time series features include: yield dynamic change rate, reaction temperature change rate, reaction pressure change rate, hydrogen-oil ratio change rate, and temperature-pressure combined change rate.

[0043] Preferably:

[0044] 1) Temperature-pressure interaction effect = reaction temperature × reaction pressure

[0045] 2) The hydrogen-oil ratio adjustment factor aver_ho_ratio is obtained by the following formula:

[0046]

[0047] In the formula, ho_ratio represents the hydrogen-oil ratio, and aver_ho_ratio represents the average hydrogen-oil ratio;

[0048] 3) The catalytic efficiency index cata_eff_index is obtained by the following formula:

[0049]

[0050] In the formula, product_weight represents the product output during the period when the catalyst is used, cata_age represents the catalyst usage duration, and cata_act represents the catalyst activity index.

[0051] Preferably:

[0052] 1) The yield dynamic change rate yield_change_rate is obtained by the following formula:

[0053]

[0054] In the formula, yield represents the yield of pyrolysis gasoline, t represents any moment, and Δt represents the unit time;

[0055] 2) The reaction temperature change rate reac_temp_rate is obtained by the following formula:

[0056]

[0057] In the formula, reac_temp(t) represents the temperature;

[0058] 3) The reaction pressure change rate reac_press_rate is obtained by the following formula:

[0059]

[0060] In the formula, reac_press(t) represents the pressure;

[0061] 4) The hydrogen-oil ratio change rate ho_ratio_rate is obtained by the following formula:

[0062]

[0063] In the formula, ho_ratio(t) represents the hydrogen-oil ratio;

[0064] 5) The temperature-pressure combined change rate temp_press_change_rate is obtained by the following formula:

[0065]

[0066] Advantages of the present invention

[0067] (1) The fitting ability for mutation points is enhanced. When the yield mutates, the prediction error is small, enhancing the practicability of the model.

[0068] (2) Sufficient feature engineering. Basic features are constructed from four directions: main components, operations, catalysts, and the environment; then composite features are constructed using some basic features; finally, time series features are constructed for the strong correlation between yield and time. Thus, a feature library for predicting the yield of pyrolysis gasoline is constructed.

[0069] (3) The data distribution is balanced. By using the SMOT algorithm, the proportion of mutation point data in the training set is increased to prevent the model from tending to fit the mainstream data distribution, ensuring an enhanced prediction ability of the model for mutation points. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Flow chart of training the pyrolysis gasoline hydrogenation yield prediction model of the present invention.

[0071] Figure 2 Real-time prediction flow chart of the pyrolysis gasoline hydrogenation yield prediction of the present invention.

[0072] Figure 3 Schematic diagram of one of the processes for synthesizing samples in the embodiment.

[0073] Figure 4 Schematic diagram of the second process for synthesizing samples in the embodiment.

[0074] Figure 5 Schematic diagram of the third process for synthesizing samples in the embodiment.

[0075] Figure 6 Schematic diagram of the fourth process for synthesizing samples in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0076] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0077] Combined with Figure 1 and Figure 2 , a method for predicting mutations in the pyrolysis gasoline hydrogenation yield considering temporal characteristics, which includes the following steps:

[0078] S1. Collect real-time data of pyrolysis gasoline hydrogenation:

[0079] S2. Input the real-time data into the pyrolysis gasoline hydrogenation yield mutation prediction model;

[0080] S3. The model outputs the yield prediction result.

[0081] Over time, the data distribution may change (this situation is called "concept drift"), which means that the data pattern when initially training the model may no longer be applicable to the current or future data. Therefore, it is necessary to retrain the model at a certain period (such as daily, weekly) using historical data in the recent period (such as one month, half a year).

[0082] In Figure 1 , the construction of the feature engineering library is specifically as follows

[0083] The existing characteristics, including olefin content, aromatic content, naphthene content, paraffin content, and other impurity content, are all the main components of pyrolysis gasoline and cannot fully describe the yield prediction, especially the reasons for the yield mutation. Therefore, an attempt is made to add more characteristics, especially to conduct in-depth analysis on the dynamic changes of certain operating conditions.

[0084] Build basic characteristics from four directions: main components, operation, catalyst, and environment; then use some of the basic characteristics to build composite characteristics; finally, construct time series characteristics based on the strong correlation between yield and time. Thus, a feature library for pyrolysis gasoline yield prediction is constructed, as shown in Table 1:

[0085] Table 1 Feature Library

[0086]

[0087] The specific description is as follows:

[0088] (1) Main component characteristics

[0089] 1) Olefin content ole_per: The mass percentage of olefins in the processing of pyrolysis gasoline, unit: %

[0090] 2) Aromatic content arom_per: The mass percentage of aromatics in the processing of pyrolysis gasoline, unit: %

[0091] 3) Naphthene content nap_per: The mass percentage of naphthenes in the processing of pyrolysis gasoline, unit: %

[0092] 4) Paraffin content par_per: The mass percentage of paraffins in the processing of pyrolysis gasoline, unit: %

[0093] 5) Impurity content impur_per: The mass percentage of impurities in the processing of pyrolysis gasoline, unit: %

[0094] (2) Operation characteristics

[0095] 1) Reaction temperature reac_temp: The actual working temperature in the reactor, which is an important parameter affecting the chemical reaction rate and selectivity, unit: °C

[0096] 2) Reaction pressure reac_press: The pressure inside the reactor, unit: Mpa

[0097] 3) Reaction time reac_time: The residence time of the material in the reactor, that is, the time length from the material entering the reactor to leaving the reactor, unit: h

[0098] 4) Hydrogen-oil ratio ho_ratio: It refers to the mass ratio of hydrogen to pyrolysis gasoline, reflecting the ratio of hydrogen to the raw material, dimensionless

[0099] (3) Catalyst Characteristics

[0100] 1) Catalyst Activity Index cata_act: An index that measures the ability of a catalyst to promote a specific chemical reaction, dimensionless

[0101] 2) Catalyst Usage Duration cata_age: The total time that the catalyst has been used cumulatively since its last regeneration, unit h

[0102] 3) Catalyst Regeneration Times cata_regen: The number of times the catalyst has undergone regeneration treatment, dimensionless count

[0103] (4) Environmental Characteristics

[0104] 1) Environmental Temperature env_temp: The air temperature outside the production workshop or factory, unit °C

[0105] 2) Atmospheric Humidity rela_humi: The content of water vapor in the air, usually expressed as relative humidity, unit %

[0106] (5) Composite Characteristics

[0107] In machine learning, constructing composite characteristics means creating new features from existing original features to capture complex relationships and patterns in the data. Doing so can significantly improve the performance of the model, especially when dealing with complex datasets. Constructing composite characteristics has important significance, such as extracting hidden information, strengthening professional domain knowledge, enhancing model performance, etc.

[0108] 1) Temperature-Pressure Interaction Effect temp_press_inter

[0109] The temperature-pressure interaction effect = reaction temperature × reaction pressure, that is

[0110] temp_press_inter = reac_temp × reac_press

[0111] This feature can help capture the complex mutual influence between temperature and pressure.

[0112] 2) Hydrogen-Oil Ratio Adjustment Factor adj_ho_ratio

[0113] The average hydrogen-oil ratio over a past period (such as one week, one month) is aver_ho_ratio, and the hydrogen-oil ratio adjustment factor

[0114] is

[0115]

[0116] Using the ratio relative to the average value to standardize the hydrogen-oil ratio helps identify situations that deviate from normal operation.

[0117] 3) Catalytic efficiency index cata_eff_index

[0118] The product output during the period when the catalyst is used is product_weight, and the catalytic efficiency index is

[0119]

[0120] Evaluate the current catalytic efficiency by combining the usage duration and activity index of the catalyst.

[0121] (6) Temporal characteristics

[0122] Introduce a time variable:

[0123] At any time t and per unit time Δt. Characterize the temporal characteristics through the change rate of some characteristics within the unit time Δt. For any characteristic f, f(t) represents the value of the characteristic at time t, and f(t + Δt) represents the value of the characteristic at time t + Δt.

[0124] 1) Yield dynamic change rate yield_change_rate

[0125] The yield of pyrolysis gasoline is yield, and the yield dynamic change rate is

[0126]

[0127] Calculate the change rate of the yield within a specific time period to help identify rapidly changing trends.

[0128] 2) Reaction temperature change rate reac_temp_rate

[0129] The reaction temperature change rate represents the change speed of the temperature within a certain time interval, that is

[0130]

[0131] This helps to capture the influence of temperature fluctuations on the pyrolysis gasoline hydrogenation process.

[0132] 3) Reaction pressure change rate reac_press_rate

[0133] The reaction pressure change rate represents the change speed of the pressure within a certain time interval, that is

[0134]

[0135] Pressure changes may significantly affect the chemical reaction rate and product distribution.

[0136] 4) Hydrogen-oil ratio change rate ho_ratio_rate

[0137] The change rate of hydrogen-oil ratio represents the change speed of the hydrogen-oil ratio within a certain time interval, that is

[0138]

[0139] The hydrogen-oil ratio is crucial for controlling the hydrogenation reaction.

[0140] 5) The combined change rate of temperature and pressure temp_press_change_rate

[0141] The combined change rate of temperature and pressure is used to capture the impact of the simultaneous changes in temperature and pressure on the system, that is

[0142]

[0143] This composite feature can help the model better understand the dynamic changes under the combined action of multiple factors.

[0144] In Figure 1 the following are the related descriptions of SMOTE generating mutant samples:

[0145] In the prediction of the hydrogenation yield of pyrolysis gasoline, there are a small number of (usually less than 5%) yield mutant samples. The number of mutant samples is much less than that of normal samples, resulting in a class imbalance problem. At this time, the normal data accounts for the absolute majority, and the prediction training process will be dominated by normal data prediction, thus causing inaccurate prediction of mutant data.

[0146] To address the class imbalance problem, the SMOTE algorithm can be used to oversample the minority class samples (i.e., mutant samples). SMOTE increases the number of minority class samples by generating synthetic samples between minority class samples, thereby balancing the class distribution in the dataset. This method helps improve the model's learning ability for minority class samples and thus improves the prediction effect for mutant situations. The specific steps are as follows:

[0147] (1) Set SMOTE parameters

[0148] Set the relevant parameters of SMOTE according to specific requirements. For example, the number of nearest neighbors (k_neighbors) can be specified, as well as whether to randomly shuffle the sample order (random_state);

[0149] (2) Perform oversampling

[0150] Apply SMOTE to the minority class samples in the training set to generate synthetic samples. This process involves the following steps:

[0151] 1) For each minority class sample, find its k nearest neighbor samples. For example Figure 3As shown, where solid circles represent the majority of samples, i.e., normal data; 3 solid triangles represent the minority class samples, i.e., mutant data. Select k = 2, and all mutant data find their two nearest neighborhood samples respectively, as Figure 4 shown by the 6 hollow triangles in

[0152] 2) Randomly select a nearest neighbor sample and generate a new synthetic sample between the two. As Figure 5 shown by the 6 solid triangles in

[0153] 3) Repeat the above process until the required number of samples is reached. As Figure 6 belonging to the 9 solid triangles in

[0154] In the above example, the initial number of mutant samples is 3, the number of normal samples is 28, and the proportion of mutant samples is 3 / 31 = 9.7%, and the sample quantity is extremely unbalanced. After being processed by the SMOTE algorithm, the number of mutant samples is 9, and the proportion of mutant samples is 9 / 37 = 24.3%, solving the problem of sample imbalance.

[0155] (3) Conduct quality inspection on the synthetic samples

[0156] Although SMOTE can effectively increase the number of minority class samples, the quality of the generated synthetic samples also needs to be verified. The distribution of the newly generated samples can be checked through visualization means (such as two-dimensional or three-dimensional scatter plots) to ensure that they are reasonably distributed around the original samples without introducing too much noise.

[0157] In Figure 1 , the specific description of the LSTM prediction algorithm is as follows:

[0158] Long Short-Term Memory (LSTM) is a special type of Recurrent Neural Network (RNN) that performs well in processing and predicting time series data. The core of LSTM is the memory cell, which can store and update long-term information. The memory cell can determine which information needs to be remembered and which information can be forgotten, thus effectively dealing with the problem of long-term dependencies. Combined with its own gating mechanism, it has advantages such as long-term dependency modeling, adaptive learning, and feature extraction ability, which is more conducive to solving the prediction problem of time series data.

[0159] Steps for prediction using LSTM:

[0160] (1) Data preprocessing

[0161] Standardization / normalization: Since LSTM is sensitive to the input scale, it is recommended to first perform standardization or normalization processing on numerical features.

[0162] Time series segmentation: Convert the data into a form suitable for LSTM input, which is a three-dimensional array (number of samples, number of time steps, number of features). For example, if you have data from the past 5 time points to predict the yield at the next time point, each sample should contain data for 5 time steps.

[0163] Create training and test sets: Divide the dataset according to a certain ratio (such as 80:20), ensuring that the distribution of mutant samples in the two sets reflects the ratio of the original dataset as much as possible.

[0164] (2) Build an LSTM model

[0165] Input layer: Define the input dimension according to the shape of your data, usually (number of time steps, number of features).

[0166] LSTM layer: Determine the number and number of layers of LSTM units.

[0167] Dense layer: Output layer. For regression tasks, there is usually only one neuron and no activation function is used.

[0168] (3) Compile the model

[0169] Select a loss function: For regression problems, the commonly used loss function is Mean Squared Error (MSE).

[0170] Optimizer: Select a suitable optimizer, such as the Adam optimizer, which can automatically adjust the learning rate according to the gradient.

[0171] Evaluation metrics: In addition to the loss function, other evaluation metrics can also be specified, such as Mean Absolute Error (MAE), etc.

[0172] (4) Train the model

[0173] Set hyperparameters: Determine hyperparameters such as batch size and number of epochs.

[0174] Training process: Use the training set to train the model and monitor the performance on the validation set to avoid overfitting.

[0175] (5) Model evaluation and tuning

[0176] Evaluate the model: Use the test set to evaluate the model performance, mainly focusing on prediction accuracy (such as MSE, MAE). Pay special attention to the model's prediction ability for mutant situations.

[0177] Hyperparameter tuning: If the initial results are not satisfactory, you can adjust hyperparameters such as the number of LSTM layers, the number of units, the dropout ratio, and the learning rate through methods such as grid search or random search.

[0178] (6) Prediction and Deployment

[0179] Predicting new data: Once the model is trained and reaches a satisfactory performance level, it can be used for actual predictions. Continuous monitoring and updating: As Figure 1 shown, retrain the model regularly to adapt to new data pattern changes and ensure the long-term effectiveness of the model.

[0180] The specific embodiments described in this document are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments or use similar means for substitution, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for predicting sudden change of pyrolysis gasoline hydrogenation yield considering time series characteristics, characterized in that It includes the following steps: S1. Collect real-time data of pyrolysis gasoline hydrogenation: S2, real-time data input cracking gasoline hydrogenation yield mutation prediction model, the model is constructed by the following steps: S2-1. Obtain historical data on hydrogenation of pyrolysis gasoline; S2-2, data preprocessing; S2-3, based on the preprocessed data, all data samples are obtained, including ① using the SMOTE algorithm to oversample minority class samples; ② merging with normal data samples to generate all data samples; ③ building a feature library; S2-4, all data samples are divided into training set and test set; S2-5. Build a long short-term memory network LSTM model, train and test it, and obtain a prediction model for the sudden change of the pyrolysis gasoline hydrogenation yield; S3. The model outputs the yield prediction results.

2. The method according to claim 1, characterized in that The model is retrained according to the historical data of the period.

3. The method according to claim 1, characterized in that The data preprocessing includes: (1) Standardization or normalization; (2) Time series segmentation, converting the data into a three-dimensional array suitable for LSTM input (number of samples, time steps, number of features); (3) Data cleaning.

4. The method according to claim 1, characterized in that The SMOTE algorithm is used to oversample the minority class samples, which includes the following steps: (1) Set SMOTE parameters; (2) Perform oversampling, including: 1) For each minority class sample, find its k nearest neighbor samples; 2) Randomly select a nearest neighbor sample and generate a new synthetic sample between the two; 3) Repeat the above process until the required number of samples is reached; (3) Perform quality check on the synthetic samples.

5. The method according to claim 1, characterized in that The feature library includes: main component features, operation features, catalyst features, environmental features, composite features, and time series features. Specifically: The main component characteristics include: olefin content, aromatic content, cycloalkane content, paraffin content, and impurity content; The operating characteristics include: reaction temperature, reaction pressure, reaction time, hydrogen-to-oil ratio; Catalyst characteristics include: catalyst activity index, catalyst usage time, catalyst regeneration times; Environmental characteristics include: ambient temperature, atmospheric humidity; Composite features include: temperature-pressure interaction effect, hydrogen-to-oil ratio adjustment factor, and catalytic efficiency index; The timing characteristics include: dynamic change rate of yield, reaction temperature change rate, reaction pressure change rate, hydrogen-oil ratio change rate, and temperature-pressure combined change rate.

6. The method according to claim 5, characterized in that: 1) Temperature-pressure interaction effect = reaction temperature × reaction pressure 2) The hydrogen-to-oil ratio adjustment factor aver_ho_ratio is obtained by the following formula: In the formula, ho_ratio represents the hydrogen-to-oil ratio, and aver_ho_ratio represents the average hydrogen-to-oil ratio; 3) The catalytic efficiency index cata_eff_index is obtained by the following formula: In the formula, product_weight represents the product output during the period of time when the catalyst is used, cata_age represents the usage time of the catalyst, and cata_act represents the activity index of the catalyst.

7. The method according to claim 5, characterized in that: 1) The yield dynamic change rate yield_change_rate is obtained by the following formula: In the formula, yield represents the yield of cracking gasoline, t represents any time, and Δt represents unit time; 2) The reaction temperature change rate reac_temp_rate is obtained by the following formula: Where, reac_temp(t) represents temperature; 3) The reaction pressure change rate reac_press_rate is obtained by the following formula: Where reac_press(t) represents pressure; 4) The hydrogen-to-oil ratio change rate ho_ratio_rate is obtained by the following formula: Where, ho_ratio(t) represents the hydrogen-to-oil ratio; 5) The combined temperature and pressure change rate temp_press_change_rate is obtained by the following formula:

Images