Capacity estimation of lithium-ion batteries for real-world electric vehicles based on multi-machine learning fusion

• The median-based Ah integration method is developed to calibrate label capacity. • 36 potential features are extracted from the real-world battery operations. • A combined CNN-GRU is fused with GPR to achieve accurate battery capacity estimation. • The proposed method provides accurate SOH estimation with an error of 2%. Accurate estimation of lithium-ion battery pack capacity under real-world driving conditions remains a significant challenge due to complex operating profiles and varying temperatures. This study proposes a multiple machine learning fusion approach, integrating convolutional and gated recurrent unit networks (CNN-GRU) with Gaussian process regression (GPR), to estimate battery pack capacity. First, the battery labeled capacity is derived using a variant of the ampere-hour integration method. Driving conditions are then extracted based on current, voltage, and temperature data from real-world electric vehicles (EVs). Moreover, the statistical values of battery charging data are analyzed, and the Pearson correlation coefficient is used to identify the optimal health features. Next, a base model combining CNN and GRU is established to capture long-term degradation trends, supplemented by a Gaussian process regression (GPR) model to compensate for temperature-induced errors. Validation using real-world EV data demonstrates exceptional performance, achieving a root mean square error of 0.87 Ah, maximum absolute error of 2.38 Ah, and average absolute error of 0.58 Ah while maintaining estimation errors within 2% of actual capacity throughout the battery's operational lifespan. These quantitative results demonstrate the method's strong potential for practical battery management system (BMS) applications.