<p>Remaining useful life (RUL) predictions contribute significantly to estimating reliability and guiding maintenance actions. However, it is difficult to extract multi-dimensional and multi-scale features for RUL prediction. This is because device degradation is affected by diverse factors, both temporal dynamics and spatial correlations, and exhibits characteristics in different dimensions and scales. Extracting multi-dimensional and multi-scale features can better capture transient anomalies and prolonged degradation patterns. Moreover, it is challenging to balance the prediction accuracy with its efficiency. Complex models achieve accurate predictions through long time and high memory, while lightweight models sacrifice accuracy for high efficiency. To overcome the above challenges, a spatio-temporal hybrid method with a parallel structure is proposed to predict RUL. First, a bidirectional temporal convolution with multiple scales is constructed to capture temporal dependencies. Second, the modified Informer with depthwise separable convolution (DSC) and pyramid self-attention is constructed to capture spatial dependencies. Third, RUL values are obtained by using fused feature mapping. The efficacy of our method is evaluated on the CMAPSS and N-CMAPSS simulation platforms, respectively. Compared to the prevalent model, the root mean square errors have been reduced by 8.45% and 6.86%, the scores have been reduced by 19.44% and 38.03%, the testing times have been reduced by 17.96% and 21.88%, and the memory has been reduced by 12.41% and 17.15%. These results show that our method has better accuracy and efficiency in RUL prediction compared to other methods.</p>