Abstract:
To address the limited ability of existing mine power load forecasting methods to jointly model sudden changes, attenuation of long-term temporal dependencies, and drift across operating conditions, a deep learning-based mine power load forecasting model was developed. Based on a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network, the model was improved in three aspects: input features, network architecture, and hyperparameters. The maximal information coefficient was introduced to adaptively select input features. A Squeeze-and-Excitation (SE) attention mechanism was introduced after the convolutional layers to adaptively weight key feature channels, enhancing the representation of local load fluctuations and sudden changes. During temporal modeling, some Long Short-Term Memory (LSTM) units were replaced with Gated Recurrent Units (GRUs), and residual connections were introduced to improve the model's ability to fit trends in mine power load variations. Using Mean Absolute Percentage Error (MAPE) as the optimization objective, Bayesian Optimization (BO) was employed to iteratively search for better hyperparameter combinations, improving the match between model parameters and the temporal characteristics of mine loads. Experimental results showed values of 19.87 kW, 25.68 kW, and 2.18% for Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and MAPE, respectively, with the coefficient of determination
R2 reaching 0.983. When sudden changes occurred at load nodes, prediction accuracy reached 95.21%, with a mean relative error of only 3.57%. During 24 h of continuous operation, the mean peak load prediction error was as low as 3.39%, and the mean time series consistency index was 0.988. The training time per epoch was 20.8 s, and the time per prediction was 118.6 ms, the balance between prediction accuracy and efficiency is achieved.