Predicting remaining useful life (RUL) of lithium batteries is crucial for predictive maintenance and health management. Existing methods suffer from issues such as high training data requirements, low computational efficiency, and difficulty meeting real-time demands due to factors like operating environments and manufacturing processes. To address this, this paper proposes a lithium battery RUL prediction method based on multi-scale feature extraction and Extreme Learning Machine (ELM) ensemble. To address the challenge of extracting spatio-temporal features from complex lithium battery lifetime data, this study employed Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and dilated convolution to extract multi-scale features, accurately capturing the modal variations in battery lifetime behavior. An ELM-integrated lithium battery RUL prediction model was built based on Elastic Net (EN) regularization, combining concepts from Broad Learning System (BLS) and ensemble learning, to improve the prediction model"s accuracy and robustness for complex sequential data and allow its use in resource-constrained scenarios.Experimental results on the Maryland University"s CALCE dataset reveal that the suggested strategy yields an 81.25% reduction in RMSE, a 60.43% decrease in MAPE, and a 0.74% improvement in R2 relative to the optimum deep learning model. Compared to the optimal standard machine learning model, training time increases by less than 1ms, while prediction time drops by 45%. This method successfully boosts model prediction accuracy and robustness, enabling lightweight models to perform better in resource-constrained circumstances.