小样本条件下智能布点代理模型及优化设计

Intelligent distribution agent model and optimization design under small sample condition

传统典型代理模型需要使用大量样本点实现高精度全局预测,对时间和计算资源的消耗很难满足工程和科学计算所需,鉴于优化设计依赖代理模型和优化算法,提出一种小样本条件下实现高效预测的智能布点模型。考虑到优化问题只需构建帕累托前沿附近(局部)的高精度代理模型,通过信赖域思想改进并实现智能布点方法,其采样空间与代理模型会根据预测精度的需求进行更新,提高了代理模型更新效率和预测精度。以小型非晶合金变压器为验证案例,在只有70组样本数据下对比了智能布点代理模型和传统典型正交实验方法,结果表明智能布点代理模型的预测精度相比传统代理模型提高了16%左右,验证了该代理模型在少样本情况下的有效建模和预测精度,并进行了优化设计。

Aiming at the problem that traditional typical surrogate models need to use a large number of sample points to achieve high-precision global prediction, and the consumption of time and computing resources is difficult to meet the needs of engineering and scientific computing, according to the optimization design depending on the surrogate model and optimization algorithm, an intelligent placement model is proposed for efficient forecasting under small sample conditions. Considering that for optimization, only a high-precision surrogate model near the Pareto frontier(local) needs to be constructed, the intelligent distribution method was improved and realized through the idea of the trust region, the sampling space and the surrogate model was updated according to the needs of the prediction accuracy, which improves the accuracy of the prediction. Taking a small amorphous alloy transformer as a verification case, the intelligent distribution surrogate model and the traditional typical orthogonal experimental method were compared under only 70 sets of sample data. The results show that the prediction accuracy of the intelligent distribution surrogate model is improved by about 16% compared with the traditional proxy model, which verifies the effective modeling and prediction accuracy of the surrogate model in the case of few samples, and conducts an optimized design.