battery-soh-lstm

Battery State of Health (SoH) Prediction - LSTM Model

Overview

LSTM-based model for predicting the State of Health (SoH) of electric vehicle batteries.

Architecture based on:

• CyclePatch-LSTM (BatteryLife, arxiv 2502.18807) - cycle-level tokenization
• EVBattery (arxiv 2201.12358) - real EV capacity estimation

Architecture

• Intra-cycle encoder: 2-layer FFN with LayerNorm (captures within-cycle patterns)
• Inter-cycle LSTM: 2-layer, 128 hidden units (captures cross-cycle degradation dynamics)
• Prediction head: 3-layer MLP with Sigmoid output

Performance (Test Set)

Input Features (8 per cycle)

1. Average discharge voltage
2. Charge capacity
3. Average temperature
4. Internal resistance
5. Charge time ratio
6. Voltage drop at end of discharge
7. Coulombic efficiency
8. Normalized cycle number

Usage

import torch
import json
from model import BatterySoHLSTM

# Load model
checkpoint = torch.load('battery_soh_lstm.pt', map_location='cpu')
config = checkpoint['config']

model = BatterySoHLSTM(
    n_features=config['n_features'],
    embed_dim=config['embed_dim'],
    intra_layers=config['intra_layers'],
    lstm_hidden=config['lstm_hidden'],
    lstm_layers=config['lstm_layers'],
    dropout=config['dropout'],
    bidirectional=config['bidirectional']
)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Predict SoH from 20 consecutive cycles of features
# Input: (batch, 20, 8) - 20 cycles, 8 features each
x = torch.randn(1, 20, 8)  # replace with real data
with torch.no_grad():
    soh = model(x)
    print(f"Predicted SoH: {soh.item():.4f}")

Training Configuration

• Optimizer: Adam (lr=1e-3, weight_decay=1e-4)
• Scheduler: ReduceLROnPlateau (factor=0.5, patience=5)
• Early stopping: patience=15
• Batch size: 64
• Data: 150 batteries, split by battery ID (no data leakage)

References

1. BatteryLife: A Comprehensive Dataset and Benchmark (arxiv 2502.18807)
2. EVBattery: Large-Scale EV Dataset (arxiv 2201.12358)
3. BatteryML: An Open-Source Platform (arxiv 2310.14714)