Double Enhanced Self Correcting Model
Double Enhanced Self Correcting(DESC) model is an Electrical Equivalent Circuit Model for lithium-ion batteries that also can be used for other battery types.
DESC is an enhanced derivative of Enhanced Self Correcting(ESC) Model from Gregor Plett.
Thus, same documentation for ESC model and DESC model article can be used as referenced to understand
code framework. After understanding the code framework and how does model work then one can develop own model.
All details, documentation and description can be found below
http://mocha-java.uccs.edu/BMS1/index.html http://mocha-java.uccs.edu/ECE5710/index.html http://mocha-java.uccs.edu/ECE5710/ECE5710-Notes02.pdf
In this framework I have developed python code framework for my battery model and this framework also includes optimisation and parameter estimation scripts for my model.
https://data.mendeley.com/datasets/nbtrgh5frn/ here I have done similar thing with simulink for my research paper. Now I moved the framework into Python. Here https://github.com/batterysim/esctoolbox-python there is a similar framework for Enhanced Self Correcting Model(ESC). I used some part of codes here but the structure is totatly diferent.
DESC model is an electric equivalent circuit model for Lithium-Ion Batteries. This model can be modified for your own model and the same framework can be used for modeling and estimation etc. Also I have done a lot of optimisation for my model (especialy with nonlinear least sqaures estimation) thus, one can use these optimisation scripts for different functions and it can be good start point.
First of all I have prepared this framework with Anaconda Spyder so I do recommend you to use it. Spyder provides matlab-like user interface it is conveninence. Except the DESC_1RC.py classs each file can be run individually without any parameter.
Run ocv.py file in the ocv_curve_fitting folder. This code provided from here https://github.com/batterysim/esctoolbox-python ocv.py uses charge and discharge test data and takes its average to obtain rough OCV(Open Circuit Voltage) curves for all test temperatures and then does basic inter/extraploation for curve fitting. That gives us good start point for further estimation. After runing this file it will generated ocvmodel.json file with model eta, Q parameters and OCV curves. Copy this file into the DESC_Model\model_files
Run DynamicProcess.py file. That starts parameter estimation for DESC model with dynamic test dataset. Since I used 1-RC branch for my model there are 8 paramerters. These are Eta, Q, M, g, R1, RC1, R0cha, R0dch respectively. Eta and Q can be directly calculated from the test data and for the rest I ran nonlinear least squares optimization from scipy.optimize.least_squares. There are many tricks and to lelarn about this library. Thus I recommend you to don't change settings at the beggining. For example tolerance values I used xtol=ftol=gtol=1e-6. This provides more aggressive optimisation but takes longer time. After parameter estimation one can run ocv estimation and then again parameter estimation with lower tolerance values(like 1e-8). That provides better estimation results.
DynamicProcess.py takes ocvmodel.json files and read test data and does parameter optimisation. After optimisation all results will be saved in the json file with name DESC1Model.json file. This file include estimated parameters but not OCV curves.
At first step we have got OCV with basic averaging and inter/extrapolation. This gives as OCV curves for each temperature in a array with 201 elements. So if we treat this arrays as parameter then we can do optimisation also on OCV curves. But this time we have 201 parameters instead of 6. I used 2 methods for OCV estimation. First I have started with nonlinear least square(NLSQ) optimisation but I it generates some ripples on the curves these has no physcial meaning. Because as OCVs nature OCV should be decreased while discharging. Thus with NLSQ I have applied linear inter/extrapolation to smooth this curves. It gives reasonable results somehow.
Then I have discovered sequantional Least Squares Programming(SLSQP) library in python which was what I needed. Since I am not so familiar with optimisation I didn't know constrained optimisation. SLSQP provides you constrained optimisation so I used constrain as OCV[n]>=OCV[n-1]. That means OCV values cannot be bigger than previous values. SLSQP doesn't require interpolation. Thus, it is more convinient and gives better results then nice OCV curves then NLSQ. How ever I kept both files for comparision.
For OCV curve fitting you can run OCVProcess_SLSQP.py(recommended) or OCVProcess_NLSQ.py. These files takes DESC1Model.json file and dynamic dataset. After fitting the result will be saved in the models folder with related name.
After first you have estimated OCV curves and parameters then one can run step 2-3 sequentially with lower tolerance values. That provide more aggressive estimation but takes longer time. For example for OCV estimation once I had to wait more than 10 hours.
DESC_Simulation.py can be used to simulate model with estimated parameters. Plot_Dynamic_Data.py can be used to plot dynamic test data Plot_OCV_Data.py can be used to plot OCV test data
