Dissolved Gas Prediction in Transformer Oil via the Hybrid PatchTST-TCN-BiLSTM Model and Animated Oat Optimization

Dissolved gas analysis (DGA) is a key diagnostic method for assessing transformer insulation integrity and identifying incipient faults, ensuring the reliability of power systems. However, the inherent nonlinearity, dynamic nonstationarity, and complex multivariable coupling in dissolved gas concentration data present substantial challenges for predictive models in capturing temporal dependencies and extracting high-dimensional features. To address these challenges, we propose a composite prediction model that integrates Patch time series transformer (PatchTST), temporal convolutional network (TCN), bidirectional long short-term memory (BiLSTM) networks, and the animated oat optimization (AOO) algorithm for joint hyperparameter optimization. In the proposed model, the PatchTST module encodes global temporal structures, the TCN module enhances local dynamic feature extraction, and the BiLSTM module captures complex temporal dependencies. Additionally, the AOO algorithm is used to tune the model’s architecture and training hyperparameters. Experimental validation on DGA data from a 500-kV substation in China demonstrates the effectiveness of each component in the proposed hybrid model. The proposed method achieves a MAPE of 2.31% and an R² of 0.987 on the test set, outperforming LSTM, GRU, Transformer, and SVM baselines.