电力变压器作为电力系统中的重要设备，一旦发生故障，将直接影响电力系统的正常 运行。为了避免变压器故障对电网造成影响，需要对变压器进行定期检测和维修。目前， 常采用溶解气体分析法对变压器进行故障诊断，但存在诊断准确率偏低的问题。本文通过 分析变压器油中溶解气体成因、变压器故障类型和特征气体类型，以变压器油中特征气体 含量为数据基础，提出了一种基于特征提取和改进野犬算法优化深度混合核极限学习机 （Deep Hybrid Kernel Extreme Learning Machine，DHKELM）的变压器故障诊断模型。主 要研究内容如下： 针对变压器油中溶解气体数据单一和常规故障诊断方法准确率低的问题，提出了一种 比值法和核主成分分析相结合的特征提取方法以及改进极限学习机的变压器故障诊断模 型。首先通过比值法构造出15维气体比值数据样本，并通过对比不同核函数下累计贡献 率，确定采用高斯核主成分分析对15维气体比值数据样本进行特征提取。为了提高极限 学习机的故障诊断正确率，通过分析不同核函数优势机制，构造了一种多项式核函数与高 斯核函数加权组合的多核函数，并引入自动编码器对极限学习机进行改进，建立深度混合 核极限学习机模型。最后采用不同故障诊断模型对降维前后的数据进行故障诊断，实验结 果表明，本文所提出的结合核主成分分析和深度混合核极限学习机故障诊断模型的正确率 达90%，较改进前提高了11.11%。 针对深度混合核极限学习机模型诊断性能受参数取值影响的问题，提出了基于改进野 犬算法优化深度混合核极限学习机的变压器故障诊断方法。首先采用Tent混沌映射对野犬 种群进行初始化，提升种群多样性，并引入差分进化算法来增强算法的全局搜索能力，将 柯西变异和反向学习融入野犬优化算法中，通过选择概率来确定最优解，以增强算法跳出 局部最优的能力，提高其收敛速度和精度。通过引入单峰和多峰测试函数并结合评价指标， 对改进野犬优化算法的性能进行测试和分析，验证了所提改进优化算法具有更强的稳定性 和寻优能力。最后以特征处理数据集作为输入，分别建立不同寻优算法确定深度混合核极 限学习机关键参数的变压器故障诊断模型，并结合故障诊断误差图进行案例分析。实验结 果表明，相较于其他四种算法优化深度混合核极限学习机的变压器故障诊断模型，采用改 进野犬优化算法优化的深度混合核极限学习机模型具有更高的故障诊断精度，其诊断正确 率达96.11%，较参数优化前提高了6.11%。

Power transformer as an important equipment in the power system, once a fault occurs, it will directly affect the normal operation of the power system. In order to avoid the impact of transformer faults on the power grid, it is necessary to carry out regular detection and maintenance of the transformer. At present, the dissolved gas analysis method is usually used for fault diagnosis of transformers, but there is a problem of low diagnosis accuracy. By analyzing the causes of dissolved gas in transformer oil, transformer fault type and characteristic gas type. Based on the characteristic gas content in transformer oil, a transformer fault diagnosis model based on the feature extraction and improved dingo optimization algorithm optimized deep hybrid kernel extreme learning machine is proposed. The main research contents of this paper are described as follows: Aiming at the problem of single data of dissolved gases in transformer oil and low accuracy of conventional fault diagnosis methods, a feature extraction method combining ratio method and kernel principal component analysis and an improved transformer fault diagnosis model of the extreme learning machine. First, the 15-dimensional gas ratio data sample is constructed by the ratio method, and by comparing the cumulative contribution rate under different kernel functions, the 15-dimensional gas ratio data sample is extracted by Gaussian kernel principal component analysis. In order to improve the fault diagnosis accuracy of the extreme learning machine, a multi-kernel function with a weighted combination of polynomial kernel function and Gaussian kernel function is constructed by analyzing the advantage mechanism of different kernel functions, and an automatic encoder is introduced to improve the extreme learning machine and establish a deep hybrid kernel extreme learning machine model. Finally, different fault diagnosis models are used to diagnose the data before and after dimension reduction. The experimental results show that the accuracy rate of the fault diagnosis model combining kernel principal component analysis and deep hybrid kernel extreme learning machine is 90%, which is 11.11% higher than before the improvement

For the problem that the diagnostic performance of the deep hybrid kernel extreme learning machine model is affected by the parameter value, this paper proposes a transformer fault diagnosis methord based on feature extraction and improved dingo optimization algorithm optimized deep hybrid kernel extreme learning machine. Firstly, utilizing Tent chaos mapping to initialize dingo population, improve population diversity, and introduce the differential evolution algorithm to enhance the global search ability, the west variation and reverse learning into dingo optimization algorithm, by selecting probability to determine the optimal solution, to enhance the ability of the algorithm out of the local optimal, improve the convergence speed and accuracy. By introducing unimodal and multimodal test functions and combining the evaluation index, the performance of the improved dingo optimization algorithm is tested and analyzed, and the proposed improved optimization algorithm has stronger stability and optimization ability. Finally, with the feature processing data set as input, the transformer fault diagnosis model with different optimization algorithms to determine the key parameters of the deep hybrid kernel limit learning machine is established respectively, combined with fault diagnosis error diagram, the case analysis is made. The experimental results show that compared with the transformer fault diagnosis model of the deep hybrid kernel limit learning machine, the deep hybrid kernel extreme learning machine model optimized by the improved dingo optimization algorithm has higher fault diagnosis accuracy, and the diagnosis accuracy reaches 96.11%, which is 6.11% higher than before parameter optimization.