Abstract:

To address the insufficient accuracy of abnormal-state identification in high-rise buildings distribution boxes, an improved autoencoder model with fusion prediction is proposed. First, the data were normalized using the min-max method to eliminate dimensional differences. Second, the model employed a BiGRU to capture temporal dependencies in both forward and backward directions. A scaled dot-product attention (SDA) mechanism was applied to the hidden states of BiGRU to enhance key time-step information Convolutional layers and a channel attention mechanism were also incorporated to extract local correlations among features. Then, a symmetric encoder-decoder structure was adopted, and hidden states from both encoding and decoding phases were concatenated to form a "reconstruction-prediction" dual-task framework, enabling the prediction of electrical parameters for the next time step. Finally, a scoring function based on prediction error was used to quantify sequence anomalies. The results show that root mean square error(RMSE),mean absolute error(MAE), and coefficient of determination R² are 0.000 9, 0.017 8, and 0.731 3, respectively. Compared with XGBoost, gated recurrent unit(GRU), and Transformer, the model shows improved performance in MAE and R² and produces corresponding anomaly-score responses to gradual residual-current drift, abnormal temperature rise, and sudden current disturbances.

Key words: fusion prediction, high-rise buildings, autoencoder, electrical fire warning, bidirectional gated recurrent unit (BiGRU), attention mechanism