基于组合蚁群算法优化神经网络诊断变压器潜伏性故障

Power transformer incipient fault diagnosis based on neural network optimized by combined ant colony optimization

针对变压器故障诊断中BP神经网络存在收敛速度慢、易陷入局部极小值等问题,提出组合蚁群算法(Combined ACO,CACO)以优化BP神经网络的权值和阈值,加快神经网络的收敛速度并实现全局最优。该方法将带精英策略的蚂蚁系统与最大-最小蚂蚁系统进行结合,对精英蚂蚁信息素轨迹量值域范围进行区间限制,有效地解决了各信息素轨迹之间差异过大的问题,避免了局部最优和早熟收敛,提高了算法的全局搜索能力,克服了常规BP算法训练神经网络在变压器故障诊断中存在的不足。实例验证表明,CACO神经网络比BP神经网络减少了96%的迭代次数,并且在故障诊断方面,CACO神经网络的准确率达到了93.9%,远远高于BP神经网络的78.5%。

In order to solve the problems of BP neural network in transformer fault diagnosis, such as slow convergence speed and easy to fall into local minimum, combined ant colony algorithm (combined ACO, CACO) is proposed to optimize the weights and thresholds of BP neural network, to speed up convergence speed and to achieve global optimal. This method combines the elitist strategy ant system with the max-min ant system to limit the range of the pheromone trajectories of elitist ants. It effectively solves the problem of large discrepancy between different pheromone trajectories to avoid local optimum and premature convergence and improves the global search ability of the algorithm. The combined ant colony algorithm neural network effectively overcomes the shortcomings of BP algorithm training neural network in transformer fault diagnosis. The example verification shows that CACO neural network reduces the number of iterations by 96% than that of BP neural network and the accuracy of CACO neural network reaches 93.9% in fault diagnosis which is far higher than that of 78.5% of BP neural network, it also proves the effectiveness of the CACO algorithm.