Automated Evaluation of Refinery and Petrochemical Feedstocks Using a Combination of Historical Market Prices, Machine Learning, and Algebraic Planning Model Information

Abstract

Computer tool determines target feedstock for a refinery, process complex, or plant. The tool receives a dataset of market conditions and preprocesses the data based on properties of the plant. Using the preprocessed data and machine learning, the tool trains predictive models. Each predictive model calculates a breakeven value of a candidate feedstock for the given plant under an individual market condition. Different predictive models optimize for different market conditions. A trained predictive model is selected based on a current market condition. The tool applies the selected predictive model and determines whether a candidate feedstock is a target feedstock for the refinery under the current market condition.

What is claimed is: 
     
         1 . A method of determining a target feedstock for a process complex, the method comprising:
 determining, for a given process complex, breakeven values for a candidate feedstock under market conditions using a computer executed process complex planning module for the given process complex, the market conditions comprising of feedstock prices and product prices;   generating a set of vectors, wherein each vector comprises: (i) the determined breakeven value for the candidate feedstock under an individual market condition, (ii) the feedstock prices of the individual market condition, (iii) the product prices of the individual market condition;   determining multipliers for the feedstock prices and product prices of the set of vectors;   applying the determined multipliers and adjusting the feedstock prices and the product prices of the vectors by the determined multipliers;   dividing the set of adjusted vectors into subsets based on the determined breakeven values for the candidate feedstock;   transforming a subset of vectors onto a reduced dimensional space;   for each subset of vectors training, using a machine learning algorithm, models configured to calculate a breakeven value of the candidate feedstock for the given process complex;   selecting a trained model based on a target market condition; and   executing the selected trained model and determining if the candidate feedstock is a target feedstock for the given process complex under the target market condition.   
     
     
         2 . The method of  claim 1  wherein determining multipliers for the feedstock prices and product prices is based upon at least one of the crack spread, average product sales fraction, average feedstock optimal input fraction, and configuration of the target process complex. 
     
     
         3 . The method of  claim 1  wherein determining multipliers for the feedstock prices and product prices is based on an optimal planning solution determined by the process complex planning module. 
     
     
         4 . The method of  claim 1  wherein dividing the set of vectors into subsets is performed by a supervised machine learning algorithm. 
     
     
         5 . The method of  claim 4  wherein the supervised machine learning algorithm is a support vector machine. 
     
     
         6 . The method of  claim 1  wherein transforming the subset of vectors onto a reduced dimensional space is performed by an unsupervised machine learning algorithm. 
     
     
         7 . The method of  claim 6  wherein the unsupervised machine learning algorithm is a principal component analysis. 
     
     
         8 . The method of  claim 1  further comprising:
 generating vector clusters from the transformed subset of vectors; and 
 for each vector cluster training, using the machine learning algorithm, a model configured to calculate the breakeven value of the candidate feedstock for the given process complex. 
 
     
     
         9 . The method of  claim 8  wherein generating vector clusters is performed by a K-means clustering algorithm. 
     
     
         10 . The method of  claim 1  wherein the machine learning algorithm is an anisotropic kriging algorithm. 
     
     
         11 . The method of  claim 1  wherein the prices of feedstocks and the prices of products of the market conditions are historical prices of feedstocks and prices of products. 
     
     
         12 . The method of  claim 1  further comprising adding: (a) the breakeven value of the candidate feedstock under the target market condition, (b) product prices of the target market condition, and (c) feedstock prices of the target market condition, as a new vector of the set of vectors. 
     
     
         13 . The method of  claim 1  wherein each vector further comprises iv) a custom crack spread of the given process complex calculated by the process complex planning module and v) a standard crack spread.