Abstract: Oil-immersed transformers are the most widely used transformers, and the top oil temperature is a crucial parameter for assessing transformer operational reliability, monitoring internal insulation conditions, and estimating internal hot-spot temperatures. It holds great significance for analyzing potential transformer faults, facilitating transformer operation and maintenance, and achieving early warning of transformer failures. To enhance the prediction accuracy of transformer top oil temperature, this paper proposes a prediction method based on the improved coati optimization algorithm (ICOA) optimized bidirectional gated recurrent unit (BiGRU) for oil-immersed transformers. Firstly, the original top oil temperature data is decomposed into intrinsic mode functions (IMFs) of different frequencies using Variational Mode Decomposition (VMD), thereby reducing the prediction difficulty of nonlinear data. Subsequently, the traditional coati optimization algorithm is improved by introducing Circle chaos mapping, Levy flight strategy, and opposition-based learning, aiming to maximize the global optimization capability and efficiency of the algorithm. Finally, a BiGRU model is constructed, and its hyper parameters are optimized using the ICOA algorithm to more precisely capture the dynamic changes in the time series, significantly improving prediction accuracy. Comparative analysis of case studies demonstrates that compared to traditional prediction models, the proposed prediction model achieves higher accuracy in predicting the top oil temperature of oil-immersed transformers.