Back-calculating key unit parameters to achieve TEA targets about biosteam HOT 4 CLOSED

Yeah I think what you proposed works, I think Monte Carlo analysis is sufficient to answer my question, I just need to process/interpret the data and come up with a way to define the probability of each scenario. Many thanks!!!

from biosteam.

As for your first analysis, we could use scipy's minimization or root solving methods. They are pretty easy to use. You can pass the bounds for all parameters and solve for the values that give the minimum or the target IRR/Price, or whatever metric you pass to the objective function. Here is an example of how this could be done:

Single variable solution for cut-off conversion

from scipy.optimize import brentq
from biorefineries.cornstover.system import cornstover_sys, cornstover_tea, R301, R302

# Should return 0 where IRR = 0.15
def f(x): # objective function
    R301.saccharification[2].X = x # Set glucan to glucose conversion of R301
    cornstover_sys.simulate()
    return  0.15 - cornstover_tea.solve_IRR()
x = brentq(f, 0, 1) # Done!

Multivariate solution to maximize IRR

from scipy.optimize import minimize
from biorefineries.cornstover.model import cornstover_model
def f(x): # objective function
    # p1, p2, p3, ... = x
    # We can set variables anyway we want here.
    # We could also use BioSTEAM's Model object to do this.
    results = cornstover_model(x)
   return - results['IRR'] # We are minimizing negative IRR
solution = minimize(f, bounds, ...)
p1, p2, p3, ... = solution.x

I don't believe we should add these methods inside biosteam objects. We would still be requesting the same info (e.g. an objective function, bounds, and all the other optional arguments in scipy functions). However, once someone gets this done, we can add their code as an "example recipe" in the docs.

As for your second analysis (the Medium one). Running Monte Carlo again and again for each parameter we change would take several weeks. Wouldn't all the data required to find this already be in your first Monte Carlo analysis? I would suggest preprocessing the data, then using sklearn's SVR to build a machine learning (ML) model to perform the calculations (to vary whatever parameters you need). We can add a method in biosteam to preprocess data and return a ML model, for those people who are timid about ML.

As for the last analysis (the Hardest one). I don't think this is all that hard. You have the distributions of all parameters in the Model object, and each chaospy Distribution object has a method to get cumulative distribution. All Monte Carlo results are saved in a data frame, just need to pass these values to each parameter.

from biosteam.

Thanks for the explanations and helpful links, I'll give the last one a try and see how that goes.

My proposed approaches aside, what would you do if you were to solve the question I tried to address - what unit/process assumptions are needed to meet a certain TEA target? Are there simpler or more robust ways? Thanks!

from biosteam.

Hmm, if what we want is the worst each parameter can perform to meet a certain TEA target. Then, after we know the relationship between all parameters and the target (e.g. through Spearman coefficients), we could brute test a grid of parameter values that matter the most. You can make a grid of values using numpy, pass it to the model using the load_samples method, evaluate, then filter the scenarios. This should give us all the results we need to conduct further analysis. For example, we can get all scenarios where a parameter is at its minimum

import numpy as np
# Let's say results is an array of filtered results where 
# each row is a scenario and each column is a parameter.
p0s = results[:, 0]
p0_min = p0s.min() # minimum possible value for parameter 0
index, = np.where(p0s ==p0_min)
scenarios_with_p0_min = results[index]

This is just one analysis you could do. But the grid should give you all the data you need.

from biosteam.