Battery State-of-Health Estimation Using the Hybrid Machine Learning Method Optimized by a Triangulation Topology Aggregation Optimizer with Partial Constant-Current Charging Data

Abstract Lithium-ion batteries have found extensive application in electric vehicles as well as large-scale energy storage systems, among which the accurate estimation of battery state-of-health (SOH) plays a critical role in tracking capacity degradation, reflecting aging characteristics, and guaranteeing the operational safety of battery systems. However, most of the existing estimation methods tend to depend on sophisticated algorithm frameworks as well as full-range charging data under standard operating conditions, hindering their deployment in an embedded battery management system. In this paper, a new SOH hybrid machine learning method is proposed using a multiple support vector regression-bidirectional long short-term memory (MSVR-BiLSTM) network, further optimized by adopting the triangulation topology aggregation optimizer (TTAO) with partial constant-current charging data. Only two charging health features are selected from the partial-constant-current charging stage. A multiple support vector regression (MSVR) framework with linear, polynomial, and Gaussian kernels is constructed for preliminary SOH estimation, and its hyperparameters are optimized by TTAO for better adaptability and stability. The outputs of these SVR models are then fed into a BiLSTM network to further improve estimation accuracy. The feasibility and superiority of the proposed method are verified based on the public NASA, CALCE, and Oxford data sets. The test results show that MSVR-BiLSTM has the highest accuracy, with a minimum MAE of 0.09%, an MAPE of 0.10%, an RMSE of 0.11%, and a maximum R2 of 0.9997, verifying its effectiveness and advantages.