Abstract: In the context of the global response to environmental pollution and climate change, countries have begun to pay attention to energy system reform and economic development to ensure low carbon transition. Among them, the development of low carbon transportation has become an important aspect of green transportation system construction. The development of electric vehicle technology can effectively reduce energy consumption and environmental pollution. However, with the recent reports of new energy vehicle safety accidents at home and abroad, the safety of lithium-ion batteries has attracted increasing attention from the industry. To prevent overcharging and overdischarging from affecting battery life and safety during use, a complete battery management system is required to control and manage a lithium-ion battery. The state of charge (SOC) used to reflect the remaining capacity of a battery is one of the key parameters. Therefore, an accurate SOC value is of significance to the safety of lithium-ion battery use and the safety performance of new energy vehicles. The low online estimation accuracy of the SOC of lithium-ion batteries and the estimation accuracy of the equivalent circuit model method are inconsistent with the model complexity. This study improved the extended Kalman filtering (EKF) algorithm and established a SOC estimation error prediction model based on the extreme learning machine (ELM) algorithm, which used the operating voltage and current of the battery as input and the SOC estimation error of the equivalent circuit model method as the output. On the basis of the physical data fusion method and the error prediction model, the online estimation model of the lithium-ion battery SOC based on the equivalent circuit model method combined with the ELM was established. The simulation results showed that the improved EKF algorithm enhances the estimation precision of the algorithm. Moreover, the physical data fusion model reduces the estimation error introduced by voltage and current measurements, overcomes the contradiction between the estimation accuracy and complexity of the equivalent circuit model method, improves the estimation accuracy of the SOC, and meets the application requirement that the estimation error must be less than 5%.