Due to the lack of multiscale feature extraction and bidirectional feature learning abilities, the existing deep state-of-charge (SOC) estimators are difficult to capture: (1) localized invariant characteristics hidden in battery measurement perturbations at multiple scales; and (2) intercorrelations among measurements in both time and reverse-time orders. If these situations are not considered, it will lead to large fluctuations in estimated SOC and accumulation of estimated errors at continuous time steps To solve these problems, an estimator combining multichannel convolutional and bidirectional recurrent neural networks (MCNN-BRNN) is proposed for SOC estimation. Specifically, MCNN can extract multiscale local robust features that are invariant to perturbations from measurements on different input paths reducing the estimation fluctuations and enhancing the robustness of the estimator. Moreover, a global convolutional layer is designed to learn the intercorrelations of multiscale features and preserve their temporal coherence. By this means, BRNN can capture the effective time-varying information of intercorrelated features in the forward and reverse directions to sequentially estimate SOC, thus alleviating the error accumulation and improving the overall estimation accuracy. Experiments results reveal that MCNN-BRNN outperforms the state-of-the-art estimators in terms of robustness and accuracy under the situations where multiscale perturbations and their comovements exist in measurement