Analysis and Prediction of the SOFC Parameters by Data-Driven Modeling

This repository contains the source materials and calculation results used in the study Predictive modeling of the dependence of solid oxide fuel cell cogeneration system voltage on operating parameters

In this version:

1. drop combine similar feature in SOFC_analisys.ipynb (transition from 25 to 29 features)
2. switching to a new engine (CatBoost)
3. standardization has been replaced by normalization (MinMaxScaler)
4. added SHAP force & dependency plot for XGBRandom_Forest
5. scale selection between MLP & DTs (GB & RF)
6. added example DT-graph for XGBRF
7. added example MLP-graph
8. translation of figures into Russian

fit result

model (components)	R2_score	MSE	MAE	MAPE	second
Gradient Boosting
CatBoost_default(full)	0.99725	0.9069	0.238	2.78%	2.571273
CatBoost+PCA(11)	0.99555	1.4629	0.346	4.93%	2.184407
CatBoost+SHAP(8)	0.99489	1.6808	0.379	2.89%	1.937381
CatBoost+PI(10)	0.99732	0.8843	0.259	2.12%	1.929342
CatBoost+MDI(14)	0.99692	1.0168	0.246	2.60%	2.189198
Random Forest
XGBRF_default(full)	0.99701	0.9835	0.261	1.10%	3.643324
XGBRF+PCA(29)	0.99607	1.2899	0.284	1.34%	2.862980
XGBRF+SHAP(15)	0.99696	1.0014	0.249	1.12%	1.651798
XGBRF+PI(15)	0.99668	1.0934	0.242	1.18%	1.725580
XGBRF+MDI(12)	0.99501	1.6398	0.328	1.28%	1.493196
MultiLayer Preceptron
MLP_default(full)	0.99154	2.7549	0.884	7.88%	1.854000
MLP+PCA(21)	0.99567	1.4255	0.392	3.32%	3.657471
MLP+PI(13)	0.98884	3.6077	0.978	8.80%	1.639187

Conclusion:

• full feature set with default XGBRandom_Forest model demonstrated the best results for MAE, MAPE metrics
• full feature set with default CatBoost model demonstrated the best results for time, R2 and MSE metrics
• full feature set with default MLP model demonstrated the best results for time metrics
• MLP+PCA works even slower than the default MLP, although it gives a more accurate result