The limited nature of short series presents difficulties for classical prediction models, as each may only contain partial information about the underlying pattern. A straightforward solution would be to concatenate these short series into longer ones in order to enhance the model performance. However, this approach can result in significant deviations from the original series, owing to substantial value differences at the junctions. Direct concatenation of the short series will ultimately disrupt the periodicity and regularity of the nearby data. This paper proposes a multi-series prediction model that is based on connecting series (hereafter referred to as connectors) and an Empirical Mode Decomposition (EMD) process. It normalizes the dataset first, and then two types of connectors are introduced between the short series for concatenation: the linear interpolation connector and the linear interpolation with random vibration (LRV) connector. Subsequently, the connected long series is decomposed into sub-sequences using EMD. Experimental results show that with the connectors EMD decomposes sub-sequences that are better aligned with the characteristics of the original short series. Specifically, the LRV connectors are suitable for multi-series with periodic characteristics, while linear interpolation connectors are more appropriate for the series short of such periodicity.