Optimization methods typically search for a descent direction based on the current solution. However

they do not acquire any prior knowledge that could be used to solve new problems in the future

resulting in reduced efficiency over time. This paper takes the computationally intensive problem of market value distribution and settlement for district integrated energy systems (DIES) as an example to address this issue. A method is proposed that utilizes historical information obtained during this optimization process for cognitive learning to update prior knowledge. Based on the current prior knowledge

values of current variables are judged

which can help accelerate the optimization of variable-structure subproblems. As the master problem is being solved

the probability of making a variable basis is consistently updated by learning from reduced-cost offset. In the variable-structure sub-problems

the scale of the basic variables is reduced based on this probability. Additionally

the entering basis is further selected from the basic variables with the largest decrease in space according to prior knowledge

which leads to a more effective pivoting and a quicker solution to the variable structure problems. The case study of multi-agent DIES variable-structure optimization indicates that compared to direct optimization

this method can reduce the solution time to an order of 10−1~10−3 seconds. By a robustness test and running the case on a GPU

the effectiveness

robustness

and scalability of the proposed method have been verified.