This is repository with very raw experiments for Steam Optimization and Other Oddities challenge which took 8th place (Kirill D) in Final Leaderboard. If you want to get some data, please, contact me.
Install virtual environment for python 3.8.
pip install -r requirements.txt
Finaly solution is in /experiments/exp_1.8/train_stacking_exp1_8_1.py and /experiments/exp_1.8/get_submission_ad_stacking.py scripts.
Try predict OIL desaturation using next features:
'DIP', 'AVG_ORIG_OIL_SAT', 'ORIG_OIL_H', 'TOTAL_INJ', 'TOTAL_GNTL_INJ',
'Lin_Dist_Inj_Factor', 'SGMT_CUM_STM_INJ_1', 'FT_DIST_PAT_1', 'SGMT_CUM_STM_INJ_2',
'FT_DIST_PAT_2', 'SGMT_CUM_STM_INJ_3', 'FT_DIST_PAT_3', 'TOTAL_PROD', 'Lin_Dist_Prod_Factor'
Dont use SAND and DATE features. Traing RandomTree model with next params:
n_trees = 3000
criterion = 'squared_error'
Result:
RMSE(Train, test_size=0.2) - 0.2529
RMSE(LB, test_size=0.2) - 0.3478
RMSE(LB, all data) - 0.346
All data the same as in 1.1, but train 4 gradient boosting models(by 4 fold cross validation) and predict
target by 4 models
Training gradient boosting model with next params:
n_trees = 400
lr = 0.1
Result:
RMSE(Train CV4) - 0.2423
RMSE(LB) - 0.3512
All data the same as in 1.1(but normalize features), train 4 linear regression
models(by 4 fold cross validation) and predict target by 4 models.
Training gradient LinReg model with next params:
lr = 0.1
Result:
RMSE(Train CV4) - 0.3699
RMSE(LB) - 0.383
All data the same as in 1.1 but add additional features:
1. Number of steam injectors(if SGMT_CUM_STM_INJ_N and FT_DIST_PAT_N columns is not nan,
is steam injector)
2. Move steam injectors columns in left columns side
3. Add SAND OHE feature
4. 1+2+3
Train random forest model.
Training gradient boosting model with next params:
n_trees = 3000
Result:
1:
RMSE(CV4) - 0.2517
2:
RMSE(CV4) - 0.2522
3:
RMSE(CV4) - 0.2475
4:
RMSE(CV4) - 0.2478
RMSE(LB, test_size=0.2) - 0.3398
RMSE(LB, all data) - 0.338
!!!WARNING!!! NEW CV STRATEGY - GroupKFold(group=CMPL_FAC_ID)**
The estimate was overestimated due to the fact that the model was
looking for the "oil well - target" relationship instead of the
"features - target" relationship.....
Train_Test_split without shuffle give me lower RMSE =~-0.34 on exp 1.4....
Try compute all scores on GroupKFold(cv=4), group is oil facility(CMPL_FAC_ID)
Compute scores for this models:
1. RF model from exp 1.1, results
RMSE(groupkf CV4) - 0.3469
2. GradBoost model from exp 1.2
RMSE(groupkf CV4) - 0.3519
3. LinReg model from exp 1.3
RMSE(groupkf CV4) - 0.3709
4. RF model from exp 1.4, results:
RMSE(groupkf CV4) - 0.3395
CONCLUSION - this CV strategy high correlated with LB, in next exps i will use
this CV strategy
Try time features:
plifetime_fcl, plifetime_sand_fcl, lifetime_fcl, lifetime_sand_fcl,
fcl_sand_seq_num, first_life_month
1. Tuned RF model_exp1.4 + normalized fcl_sand_seq_num
RMSE(groupKF CV4) - 0.3309
RMSE(LB, trained on all data) - 0.3241
2. Tuned RF model_exp1.4 + seq_num_un, fcl_life_time, fcl_sand_life_time
RMSE(groupKF CV4) - 0.3274
RMSE(LB, trained on all data) - 0.3244
3. As in 2, but add sands_num
RMSE(groupKF CV4) - 0.3281
RMSE(LB, trained on all data) - 0.3245
Try other method OHE of SAND:
1/4 codes of types of SAND
1/5 codes of number of SAND
Try stacking for best solution
1. 1.5exp Solution + new sand OHE features <- lost .py scripts...
RMSE(groupKF CV4) - 0.3280
2. Stacking with features preprocessing from exp. 1.5
RF_exp1.5 + SVR + LGBMRegressor, final_estimator - RidgeCV
RMSE(groupKF CV4) - 0.3258
RMSE(LB, trained on all data) - 0.3204
Add remaining features to exp1.5.2 approach:
1. ADD TOTAL_INJ, TOTAL_GNTL_INJ, TOTAL_PROD on normalized by facility/facility+sand variant
RMSE(groupKF CV4) - 0.3243
RMSE(LB, trained on all data) - 0.3222
2. ADD SGMT_CUM_STM_INJ_[1/2/3] on normalized by facility/facility+sand variant
RMSE(groupKF CV4) - 0.3257
3. ADD sand group feature B/C/E/D
RMSE(groupKF CV4) - 0.3265
4. Mean/STD by facility, facility+SAND features
RMSE(groupKF CV4) - 0.3273
5. Change of TOTAL_INj, TOTAL_GNTL_INJ, TOTAL_PROD
RMSE(groupKF CV4) - 0.3272
6. Mean ORIG_SAT and TOTAL_GNTL_INJ for each sand group
RMSE(groupKF CV4) - 0.3260
7. Time feature(year)
RMSE(groupKF CV4) - 0.3277
8. Number of sand group feature
RMSE(groupKF CV4) - 0.3265
9. 1 + 2 + 3 + 6 + 8 exps.
RMSE(groupKF CV4) - 0.3211
RMSE(LB, trained on all data) - 0.3168
10. STACKING data all the same as in 9
RMSE(groupKF CV4) - 0.32029
RMSE(LB, trained on all data) - 0.3131
Feature selection
1. Permutation importance, WARNING n_trees=300
First - bad features
Weight Feature
0 0.0008 ± 0.0012 FT_DIST_PAT_3
1 0.0001 ± 0.0001 SAND_TULS_C3
2 0.0001 ± 0.0010 Lin_Dist_Prod_Factor
3 0.0000 ± 0.0000 SAND_TULS_C1
4 0.0000 ± 0.0003 groups_num
5 0.0000 ± 0.0001 steam_inj_nums
6 0.0000 ± 0.0005 cum_stm2_fac_sand
7 -0.0000 ± 0.0001 SAND_TULS_B3
8 -0.0000 ± 0.0002 SAND_group_4
9 -0.0000 ± 0.0002 SAND_TULS_E3
10 -0.0000 ± 0.0002 SAND_TULS_B2
11 -0.0000 ± 0.0001 SAND_TULS_C4
12 -0.0000 ± 0.0002 SAND_num_4
13 -0.0000 ± 0.0001 SAND_TULS_D5
14 -0.0001 ± 0.0002 SAND_TULS_D4
15 -0.0001 ± 0.0001 SAND_TULS_B5
16 -0.0001 ± 0.0001 SAND_num_3
17 -0.0001 ± 0.0002 SAND_TULS_C5
18 -0.0001 ± 0.0004 fcl_sand_life_time
19 -0.0001 ± 0.0002 SAND_TULS_D3
20 -0.0001 ± 0.0001 SAND_TULS_B4
21 -0.0001 ± 0.0002 SAND_TULS_C2
22 -0.0001 ± 0.0007 DIP
23 -0.0002 ± 0.0002 SAND_TULS_D2
24 -0.0002 ± 0.0003 SAND_group_3
25 -0.0002 ± 0.0002 SAND_num_2
26 -0.0002 ± 0.0003 SAND_TULS_B1
27 -0.0002 ± 0.0003 SAND_TULS_E1
28 -0.0003 ± 0.0003 SAND_group_1
29 -0.0003 ± 0.0002 SAND_TULS_E4
30 -0.0003 ± 0.0004 total_prod_fac
31 -0.0004 ± 0.0004 SAND_group_2
32 -0.0004 ± 0.0003 SAND_TULS_D1
33 -0.0005 ± 0.0007 SGMT_CUM_STM_INJ_2
34 -0.0006 ± 0.0004 SAND_TULS_E2
35 -0.0006 ± 0.0005 SAND_num_5
36 -0.0007 ± 0.0009 FT_DIST_PAT_2
37 -0.0008 ± 0.0005 total_inj_fac
38 -0.0008 ± 0.0006 SAND_TULS_E5
39 -0.0009 ± 0.0006 SGMT_CUM_STM_INJ_1
40 -0.0012 ± 0.0011 Lin_Dist_Inj_Factor
41 -0.0012 ± 0.0007 SAND_num_1
42 -0.0013 ± 0.0008 cum_stm3_fac_sand
43 -0.0013 ± 0.0006 fcl_life_time
44 -0.0014 ± 0.0010 TOTAL_PROD
45 -0.0021 ± 0.0011 SGMT_CUM_STM_INJ_3
46 -0.0025 ± 0.0016 ORIG_OIL_H
47 -0.0026 ± 0.0014 FT_DIST_PAT_1
48 -0.0031 ± 0.0012 TOTAL_GNTL_INJ
49 -0.0063 ± 0.0018 seq_num_un
50 -0.0065 ± 0.0029 mean_group_orig_sat
51 -0.0075 ± 0.0016 cum_stm1_fac_sand
52 -0.0082 ± 0.0025 AVG_ORIG_OIL_SAT
53 -0.0111 ± 0.0032 TOTAL_INJ
54 -0.0112 ± 0.0016 total_inj_fac_sand
55 -0.0118 ± 0.0027 mean_group_sinj_gntl
2. Choice features using backward features elimination
1.1 Choised 'FT_DIST_PAT_3', 'Lin_Dist_Prod_Factor', 'groups_num', 'steam_inj_nums' features
RMSE(groupKF CV4) - 0.3189
1.2 1.1 + stacking from 1.7.10
RMSE(groupKF CV4) - 0.3173
RMSE(LB, trained on all data) - 0.3112