Transfer Learning-CNN-LSTM-Based Insulation State Prediction for Energy Storage Systems

Accurate estimation of insulation resistance, denoted as Riso, in high-voltage direct current energy storage systems plays a pivotal role in leakage protection and thermal runaway suppression. Conventional physical measurement techniques are inherently susceptible to distortions under dynamic operating conditions due to interference from parasitic capacitance. Meanwhile, emerging data-driven approaches are often bottlenecked by cross-domain distribution shifts and the scarcity of annotated full-lifecycle data. This study proposes a hybrid framework that integrates transfer learning and residual correction within a CNN-LSTM architecture, referred to as TL-CNN-LSTM + Corr. Utilizing seven-dimensional operational features as inputs, the framework employs a one-dimensional convolutional neural network to extract high-frequency transient response patterns. Simultaneously, a long short-term memory network models the long-term, non-stationary temporal evolution of insulation degradation. To circumvent systemic biases across varying scenarios, a three-stage domain adaptation strategy consisting of pre-training, freezing, and fine-tuning was developed, which is complemented by a lightweight linear residual compensator designed to rectify amplitude drifts during abrupt operational transitions. Independent evaluations using 500 sets of real-world operational data demonstrate that the proposed model achieves high-precision predictions, yielding a root mean square error of 40.595, a mean absolute error of 32.919, and an R2 value of 0.941. Furthermore, the model exhibits remarkable robustness against sensor noise and data loss. By ensuring cross-domain predictive consistency with minimal computational overhead, this framework provides a highly reliable and deployable solution for online insulation state monitoring in edge-side battery management systems.