To solve the problem of tight and complex coupling characteristics between multiple load sequences in integrated energy systems (IES) and the difficulty of accurate prediction

a short-term prediction method for IES multi-type loads based on modal decomposition and multi task learning model is proposed. First

to deal with the strong randomness of the multiple load sequences in IES

a multi-variant empirical mode decomposition and sample entropy are used to synchronously decompose and reconstruct the high-

medium-

and low-frequency modal components of the multiple load sequences. Then

a multi-task learning hybrid prediction model based on multi-head attention mechanism is established to dynamically allocate coupling characteristics. For the more complex medium to high-frequency sequences

a Transformer structure with a single encoder and multiple decoders is employed to fully capture the load fluctuation information

while for low-frequency sequences

features of steady components are extracted based on a bidirectional gated recurrent unit network. Finally

the predicted results of each component are aggregated to obtain the ultimate prediction of multiple loads. The methodology is validated using multi-type load data from Arizona State University's Tempe campus. The results show that the proposed method has an average absolute percentage error of 0.61%

0.80%

and 0.83% for electricity

cooling

and heating loads

respectively. Compared with other models

it has higher solution accuracy and computational efficiency.