光伏运维中融合KAN Transformer的TCN深度学习模型及故障检测
TCN Deep Learning Model and Fault Detection Based on Integrating KAN Transformer in Photovoltaic Operation and Maintenance

光伏发电作为清洁能源的重要组成部分，在电力系统中的集成对提高能源利用效率和减少环境污染具有重要意义。但光伏发电的波动性和不确定性给电网的稳定运行带来了挑战。本研究针对甘肃地区电力系统中光伏发电的预测问题，提出了一种创新的深度学习模型，该模型融合了Kolmogorov-Arnold Networks (KAN)、Transformer和时间卷积网络(TCN)的优势，以提高预测精度。考虑到光伏发电数据的高维性和非线性特征，KAN层被引入以提取复杂的数据模式。Transformer层通过自注意力机制有效地捕捉了时间序列中的长距离依赖关系。TCN层利用扩张卷积技术进一步增强了模型对局部时间特征的捕捉能力。为了处理数据集中的不平衡问题，本研究采用了SMOTE技术进行数据预处理，以增强模型对少数类别的识别能力。在模型训练过程中，均方误差(MSE)损失函数被用作优化目标，以最小化预测误差。此外，本研究不仅关注预测精度，还将模型应用于光伏发电故障监测，通过实时监控和分析运行数据，实现了故障的早期识别和预警。实验结果表明，所提出的模型在光伏发电预测任务上表现出色，相比于现有方法具有更高的预测精度和鲁棒性。此外，通过故障监测的应用案例，证明了模型在实际电力系统运维中的实用性。本研究的成果为光伏发电的预测和运维管理提供了新的视角和技术支持，对促进可再生能源的高效利用具有重要意义。
Photovoltaic (PV) power generation, as a vital component of clean energy, plays a significant role in integrating into the power system, enhancing energy utilization efficiency, and reducing environmental pollution. However, the volatility and uncertainty of photovoltaic power generation pose challenges to the stable operation of the power grid. This study addresses the forecasting issue of photovoltaic power generation in the Gansu region’s power system by proposing an innovative deep learning model that integrates the advantages of Kolmogorov-Arnold Networks (KAN), Transformers, and Temporal Convolutional Networks (TCN) to improve forecasting accuracy. Considering the high-dimensional and nonlinear characteristics of photovoltaic power generation data, KAN layers are introduced to extract complex data patterns. Transformer layers effectively capture long-range dependencies in time series through self-attention mechanisms. TCN layers further enhance the model’s ability to capture local temporal features using dilated convolution techniques. To address the imbalance in the dataset, this study employs the SMOTE technique for data preprocessing, enhancing the model’s recognition capability for minority classes. During model training, the mean squared error (MSE) loss function is used as the optimization objective to minimize prediction errors. Moreover, this study not only focuses on forecasting accuracy but also applies the model to photovoltaic power generation fault monitoring, achieving early identification and warning of faults through real-time monitoring and analysis of operational data. Experimental results show that the proposed model performs excellently in photovoltaic power generation forecasting tasks, with higher prediction accuracy and robustness compared to existing methods. Additionally, through the application of