To address the issue of rapid load surges in rural distribution grids during special periods such as the Spring Festival

which frequently lead to severe overload of distribution transformer

this paper proposes a research method for predicting severe overload of distribution transformer and restoring power supply based on a temporal convolutional network (TCN)-Crossformer-bi-directional long-short-term memory(BiLSTM) architecture. Firstly

a short-term load forecasting model based on TCN-Crossformer-BiLSTM is constructed. TCN extracts local time features while preserving causality. Crossformer characterizes global dependencies among external factors such as weather

dates

and holidays through cross-dimensional attention mechanisms. BiLSTM further enhances bidirectional memory and trend-fitting capabilities of sequences

thereby achieving high-precision modeling of complex non-stationary loads. Secondly

based on the predicting results and load factor discrimination criteria

severe overload risks of distribution transformer are identified

and an optimization model for power supply restoration is established with the objective of minimizing the maximum load factor during heavy and overload periods. Finally

the model is optimized and solved by combining butterfly optimization algorithm(BOA) and particle swarm optimization(PSO) approach. The case study results demonstrate that the proposed model effectively enhances the ability of load trend capture and the accuracy of perception of distribution transformer heavy overload risks.