The classification and accurate recognition of transformer operating states are very important for the stable operation of transformers and the safe power supply of power systems. At present, there are still some problems in this kind of research, such as less attention to the relevant data of transformer loads, high complexity of mechanism models and unclear correlation between oil temperature data and overload state. Therefore, this paper proposes an improved hybrid model that combines the hunter-prey optimization (HPO) algorithm with the efficient channel attention (ECA) module, applied to convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) neural networks. It is used to classify transformer operating state and identify overload fault. The data of a main transformer containing nine characteristics related to transformer load are selected as samples, and the load state category is selected through K-Means ++ clustering and transformer normal periodic load analysis. The parameters of the hybrid model were optimized by HPO to improve the performance and generalization ability. Through pre-processing and feature extraction of transformer load data, the improved model is used to recognize the load state accurately. The experimental results show that the recognition accuracy of this method can reach 99.24%, and good results are obtained in the task of transformer operating state classification.