Read automatically from the scorecard below — every claim can be checked against it.
45 % of the score comes from initial soc error (T10, 7.07 % × weight 0.100) and current sensor offset (T11, 3.77 % × weight 0.100) — improving there moves the leaderboard position most.
Accuracy holds on the never-released cell (1.49 % vs 1.63 % on the open cells) — the model generalises rather than memorising the released data.
Cold is the weak spot: 1.94 % at −20 °C against 0.72 % at 25 °C. Below zero the cell's resistance rises and the voltage curve distorts; models without temperature awareness lose the most here.
A wrong initial SOC is not corrected: 7.07 % on test 10. Pure current integration can never recover from a bad start; voltage feedback is what fixes this.
Scorecard
Average RMSE (% SOC) per test case, weighted into the leaderboard score. Lower is better.
Estimation accuracyT1All cells(reference only — every other test is a subset of it)
1.60 %
0.0000
0.0000
T2Blinded cell (m448)
1.49 %
0.1000
0.1491
T3
Key cases
The runs that separate estimators — cold and hot cycles, the blinded cell, wrong initial SOC, a biased current sensor. Each plot says why it is there.
Selected drive cycles
The thirteen cycles the original MATLAB tool plotted, one per test: blinded vs non-blinded cell (tests 1–3), drive vs charge (test 4), the three payloads on one cycle (tests 5–6), standard vs non-standard highway profiles (tests 7–8) and the temperature extremes (test 9).
m80 UDDS at -20 °C
Whole cycle · RMSE 1.59 %, max 3.1 %
Actual SOC
LSTM–Coulomb Counting (Previous Contest) - KOR
Error axis
Why this case: Test 9 — coldest condition: high resistance and a strongly non-linear OCV response — the hardest cycle for most estimators.
Wrong initial SOC (test 10)
Three cycles restarted with the estimator told the battery is at 90, 60 or 30 % while it is really full. Coulomb counting can never recover; voltage-informed estimators converge at a rate that is very visible here.
Wrong initial SOC 90 % — m80 LA92 at 25 °C
Whole cycle · RMSE 6.21 %, max 11.5 %
Actual SOC
LSTM–Coulomb Counting (Previous Contest) - KOR
Error axis
Why this case: The cycle is joined at the point where the true SOC has already fallen to 90 %, so a model that assumes a full battery at start-up begins 10 % off. What matters is how fast the estimate converges onto the truth.
Current-sensor offset (test 11)
Three cycles re-run with ±0.3 A added to the measured current. Integration-based estimators drift steadily; the plot shows whether and how fast the model corrects the bias.
Current-sensor offset -0.3 A — m1000 US06 at -10 °C
Whole cycle · RMSE 5.27 %, max 10.7 %
Actual SOC
LSTM–Coulomb Counting (Previous Contest) - KOR
Error axis
Why this case: A constant -0.3 A is added to the measured current before it reaches the model. Pure current integration drifts linearly; estimators that also use voltage should correct it.
RMSE vs. sensor offset
Test 11 — the same three cycles re-run with a constant current offset; 0 A is the unmodified run. Only ±0.3 A enters the score.
US06 at −10 °C
HWFET at 10 °C
LA92 at 40 °C
Test-case chartsThe scorecard as bar charts — tests 1–8, and RMSE against temperature (test 9).
Error by test case
Average RMS error (% SOC) for each blinded test case. Lower is better.
Error vs. ambient temperature
Test 9 — m80 cell, all blinded cycles at each chamber temperature.
All 144 cyclesEvery blinded drive cycle: per-cycle errors, and any of the plotted cycles in the time domain.
Cell
Temp.
Cycle
Duration
RMSE
MAE
MAXE
m80
-10 °C
UDDS
6.32 h
2.16
1.86
4.61
m80
-10 °C
HWFET
3.02 h
0.95
0.72
3.40
m80
-10 °C
LA92
4.74 h
2.01
1.74
4.09
m80
-10 °C
US06
2.30 h
1.15
0.94
3.53
m80
-10 °C
HWCUST
1.69 h
3.77
3.16
8.62
m80
-10 °C
HWGRADE
1.24 h
3.39
2.52
11.14
m80
-20 °C
UDDS
4.62 h
1.59
1.39
3.08
m80
-20 °C
HWFET
2.34 h
1.88
1.68
3.31
m80
-20 °C
LA92
3.54 h
1.77
1.60
3.49
m80
-20 °C
US06
1.86 h
2.19
2.01
3.91
m80
-20 °C
HWCUST
1.37 h
1.71
1.48
3.03
m80
-20 °C
HWGRADE
1.25 h
2.49
2.00
7.13
m80
0 °C
UDDS
8.75 h
1.41
1.28
2.60
m80
0 °C
HWFET
3.84 h
1.36
1.20
2.57
m80
0 °C
LA92
6.39 h
1.38
1.26
2.69
m80
0 °C
US06
2.93 h
1.75
1.61
3.29
m80
0 °C
HWCUST
2.08 h
2.48
2.31
3.80
m80
0 °C
HWGRADE
1.64 h
2.74
2.29
6.34
m80
10 °C
UDDS
11.32 h
0.45
0.42
1.04
m80
10 °C
HWFET
4.49 h
0.37
0.33
0.73
m80
10 °C
LA92
7.88 h
0.34
0.30
0.88
m80
10 °C
US06
3.42 h
0.69
0.61
1.39
m80
10 °C
HWCUST
2.42 h
1.74
1.64
2.86
m80
10 °C
HWGRADE
1.80 h
2.58
2.14
5.96
m80
25 °C
UDDS
16.13 h
0.62
0.48
1.67
m80
25 °C
HWFET
5.52 h
0.44
0.33
1.57
m80
25 °C
LA92
10.58 h
0.63
0.50
1.66
m80
25 °C
US06
4.22 h
0.46
0.35
1.62
m80
25 °C
HWCUST
2.86 h
0.65
0.63
0.94
m80
25 °C
HWGRADE
2.33 h
1.54
1.27
3.88
m80
40 °C
UDDS
12.40 h
1.99
1.63
4.04
m80
40 °C
HWFET
5.41 h
1.60
1.27
3.66
m80
40 °C
LA92
9.07 h
1.71
1.38
3.69
m80
40 °C
US06
4.17 h
1.42
1.12
3.39
m80
40 °C
HWCUST
2.95 h
0.61
0.48
1.99
m80
40 °C
HWGRADE
2.73 h
1.34
1.13
3.27
m448
-10 °C
UDDS
6.23 h
2.36
2.11
4.74
m448
-10 °C
HWFET
3.01 h
0.86
0.67
2.17
m448
-10 °C
LA92
4.59 h
1.99
1.78
3.77
m448
-10 °C
US06
2.25 h
1.04
0.89
2.24
m448
-10 °C
HWCUST
1.69 h
0.89
0.73
2.17
m448
-10 °C
HWGRADE
1.22 h
3.02
2.43
6.51
m448
-20 °C
UDDS
4.56 h
1.40
1.28
2.59
m448
-20 °C
HWFET
2.34 h
1.82
1.70
3.14
m448
-20 °C
LA92
3.46 h
1.61
1.51
2.78
m448
-20 °C
US06
1.80 h
2.16
2.03
3.92
m448
-20 °C
HWCUST
1.39 h
2.06
1.95
3.13
m448
-20 °C
HWGRADE
1.23 h
2.63
2.25
5.76
m448
0 °C
UDDS
8.60 h
1.24
1.13
2.30
m448
0 °C
HWFET
3.83 h
1.17
1.03
2.27
m448
0 °C
LA92
6.16 h
1.19
1.08
2.46
m448
0 °C
US06
2.87 h
1.69
1.57
3.12
m448
0 °C
HWCUST
2.09 h
2.38
2.29
3.50
m448
0 °C
HWGRADE
1.51 h
3.10
2.55
7.22
m448
10 °C
UDDS
11.08 h
0.45
0.40
1.09
m448
10 °C
HWFET
4.47 h
0.44
0.41
0.76
m448
10 °C
LA92
7.60 h
0.38
0.33
0.93
m448
10 °C
US06
3.35 h
0.65
0.57
1.33
m448
10 °C
HWCUST
2.42 h
1.85
1.78
2.93
m448
10 °C
HWGRADE
1.68 h
2.97
2.44
7.15
m448
25 °C
UDDS
15.59 h
0.81
0.66
1.79
m448
25 °C
HWFET
5.46 h
0.47
0.37
1.61
m448
25 °C
LA92
10.06 h
0.67
0.55
1.76
m448
25 °C
US06
4.07 h
0.47
0.39
1.28
m448
25 °C
HWCUST
2.86 h
0.62
0.60
1.02
m448
25 °C
HWGRADE
2.19 h
1.83
1.50
4.58
m448
40 °C
UDDS
12.09 h
2.13
1.76
4.28
m448
40 °C
HWFET
5.21 h
1.67
1.33
3.86
m448
40 °C
LA92
8.79 h
1.89
1.55
3.90
m448
40 °C
US06
4.06 h
1.50
1.19
3.52
m448
40 °C
HWCUST
2.92 h
0.66
0.51
2.08
m448
40 °C
HWGRADE
2.32 h
1.59
1.33
3.94
m448-N
-10 °C
UDDS
10.80 h
4.03
3.64
7.44
m448-N
-10 °C
HWFET
3.85 h
1.83
1.61
3.43
m448-N
-10 °C
LA92
6.66 h
3.30
2.97
6.34
m448-N
-10 °C
US06
2.70 h
2.16
1.99
3.71
m448-N
-10 °C
HWCUST
1.89 h
1.31
1.02
3.26
m448-N
-10 °C
HWGRADE
1.36 h
4.28
3.49
9.01
m448-N
-20 °C
UDDS
8.67 h
0.93
0.71
2.85
m448-N
-20 °C
HWFET
3.09 h
1.25
0.82
7.05
m448-N
-20 °C
LA92
5.21 h
1.67
1.40
3.50
m448-N
-20 °C
US06
2.23 h
2.60
2.40
4.51
m448-N
-20 °C
HWCUST
1.65 h
2.60
2.51
3.37
m448-N
-20 °C
HWGRADE
1.50 h
4.81
3.95
12.02
m448-N
0 °C
UDDS
12.79 h
1.68
1.53
2.85
m448-N
0 °C
HWFET
4.47 h
1.31
1.13
2.47
m448-N
0 °C
LA92
7.99 h
1.50
1.31
2.82
m448-N
0 °C
US06
3.22 h
1.86
1.70
3.49
m448-N
0 °C
HWCUST
2.31 h
2.17
2.08
3.11
m448-N
0 °C
HWGRADE
1.65 h
3.93
3.34
9.35
m448-N
10 °C
UDDS
13.92 h
0.70
0.67
1.31
m448-N
10 °C
HWFET
4.84 h
0.68
0.63
1.16
m448-N
10 °C
LA92
8.80 h
0.55
0.51
1.10
m448-N
10 °C
US06
3.56 h
0.95
0.87
1.84
m448-N
10 °C
HWCUST
2.53 h
2.11
2.02
3.47
m448-N
10 °C
HWGRADE
1.79 h
3.37
2.84
7.92
m448-N
25 °C
UDDS
15.61 h
0.50
0.40
1.33
m448-N
25 °C
HWFET
5.47 h
0.32
0.25
1.17
m448-N
25 °C
LA92
10.09 h
0.40
0.32
1.30
m448-N
25 °C
US06
4.08 h
0.45
0.37
1.24
m448-N
25 °C
HWCUST
2.86 h
0.81
0.79
1.10
m448-N
25 °C
HWGRADE
2.30 h
1.89
1.50
4.63
m448-N
40 °C
UDDS
16.75 h
1.99
1.64
3.94
m448-N
40 °C
HWFET
5.89 h
1.54
1.24
3.53
m448-N
40 °C
LA92
11.03 h
1.73
1.41
3.60
m448-N
40 °C
US06
4.44 h
1.27
1.01
3.03
m448-N
40 °C
HWCUST
3.13 h
0.57
0.47
1.77
m448-N
40 °C
HWGRADE
2.48 h
1.59
1.32
3.99
m1000
-10 °C
UDDS
5.91 h
2.41
2.16
5.05
m1000
-10 °C
HWFET
2.91 h
0.88
0.69
2.14
m1000
-10 °C
LA92
4.23 h
2.02
1.85
3.74
m1000
-10 °C
US06
2.09 h
1.08
0.93
2.30
m1000
-10 °C
HWCUST
1.67 h
1.25
1.09
3.29
m1000
-10 °C
HWGRADE
1.20 h
4.03
3.33
9.06
m1000
-20 °C
UDDS
4.31 h
1.47
1.36
2.53
m1000
-20 °C
HWFET
2.27 h
1.74
1.62
3.15
m1000
-20 °C
LA92
3.20 h
1.70
1.57
3.10
m1000
-20 °C
US06
1.74 h
1.17
1.00
2.63
m1000
-20 °C
HWCUST
1.37 h
2.93
2.80
4.10
m1000
-20 °C
HWGRADE
1.22 h
3.69
3.21
7.58
m1000
0 °C
UDDS
8.09 h
1.19
1.09
2.33
m1000
0 °C
HWFET
3.70 h
1.09
0.96
2.34
m1000
0 °C
LA92
5.60 h
1.12
1.02
2.38
m1000
0 °C
US06
2.66 h
1.62
1.51
3.14
m1000
0 °C
HWCUST
2.05 h
2.09
2.00
3.36
m1000
0 °C
HWGRADE
1.36 h
3.67
2.94
9.04
m1000
10 °C
UDDS
10.51 h
0.42
0.37
1.04
m1000
10 °C
HWFET
4.37 h
0.45
0.41
0.88
m1000
10 °C
LA92
7.06 h
0.32
0.27
0.85
m1000
10 °C
US06
3.11 h
0.58
0.49
1.38
m1000
10 °C
HWCUST
2.41 h
1.64
1.57
2.81
m1000
10 °C
HWGRADE
1.52 h
3.35
2.69
8.94
m1000
25 °C
UDDS
14.69 h
0.84
0.67
2.01
m1000
25 °C
HWFET
5.37 h
0.45
0.34
1.60
m1000
25 °C
LA92
9.30 h
0.66
0.54
1.79
m1000
25 °C
US06
3.79 h
0.47
0.39
1.36
m1000
25 °C
HWCUST
2.83 h
0.53
0.49
0.95
m1000
25 °C
HWGRADE
1.90 h
2.19
1.75
5.40
m1000
40 °C
UDDS
11.69 h
2.05
1.70
4.05
m1000
40 °C
HWFET
5.15 h
1.62
1.29
3.69
m1000
40 °C
LA92
8.33 h
1.70
1.40
3.54
m1000
40 °C
US06
3.89 h
1.25
0.98
3.03
m1000
40 °C
HWCUST
2.91 h
0.60
0.46
1.95
m1000
40 °C
HWGRADE
2.17 h
1.91
1.57
4.60
m80 UDDS at -20 °C
Whole cycle · RMSE 1.59 %, max 3.1 %
Actual SOC
LSTM–Coulomb Counting (Previous Contest) - KOR
Error axis
One hour of padded data precedes every cycle in the evaluator and is excluded from the error metrics. Charts are down-sampled for display (peaks preserved); the full 1 Hz data is in the traces download below.
Score historyEvery evaluation attempt and every change to how this submission is scored; the current score is the last row.
How a hybrid worksThe canonical structure for this model family; the author's description below gives the specific architecture.
Hybrid — standardized view
Model-based backbone + learned correction
Governing equation
SOC=SOCbackbone+gθ(I,V,T,history)
Downloads
PDF report (summary, every test case with the arithmetic, key plots, per-cycle table) · results as JSON · full 1 Hz traces of all 195 runs as a MATLAB v7 file (0.01 % SOC steps; readable with scipy.io.loadmat, see the readme variable inside).
LSTM–Coulomb Counting Hybrid SOC Estimator - Contest Model - 2023
Package
South_Korea_Updated.zip (768.4 KB) — deleted after evaluation
Visibility
Public
Evaluator
socbench-eval-0.1.0/matlab
Completed
Aug 31, 2026, 12:45 a.m. EDT
Submission ID
cmtgmygb70003la04kbq5ke0n
Non-blinded cells
1.63 %
0.1000
0.1631
Operating conditionsT4Charging
2.07 %
0.1000
0.2070
T580 kg payload
1.54 %
0.0333
0.0512
T5448 kg payload, HVAC on
1.49 %
0.0667
0.0994
T6448 kg payload, HVAC off
1.80 %
0.0667
0.1197
T51000 kg payload
1.56 %
0.0333
0.0520
T7Standard drive cycles
1.29 %
0.1000
0.1291
T8Non-standard drive cycles
2.21 %
0.1000
0.2207
T9−20 °C
1.94 %
0.0167
0.0323
T9−10 °C
2.24 %
0.0167
0.0373
T90 °C
1.85 %
0.0167
0.0309
T910 °C
1.03 %
0.0167
0.0171
T925 °C
0.72 %
0.0167
0.0121
T940 °C
1.44 %
0.0167
0.0241
RobustnessT10Initial SOC error
7.07 %
0.1000
0.7073
T11Current sensor offset
3.77 %
0.1000
0.3774
Weighted error = Σ (weight × RMSE)
2.43 %
Max error 12.0 % — largest instantaneous |estimate − truth| over every blinded cycle (on m448-N HWGRADE at -20 °C — plotted under Key cases for new evaluations); not scoredComplexity 10 ±1 · Extreme+ (informational, not scored)
State carried in z
Backbone state plus the correction model's memory.
Ahnaf Rahman · McMaster University