Enhancing Power Consumption Prediction Using Dropout-Regularized Deep Learning

Abstract

Accurate power consumption forecasting is important for energy management, smart grid operation, and efficient electricity utilization. Deep learning models can capture nonlinear relationships and temporal patterns in power consumption data, but their performance may be affected by overfitting. This study presents a comparative empirical investigation of dropout regularization across five established deep learning architectures, namely Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). The models were evaluated using three real-world power consumption datasets from Mexico, Wellington, and Ohio. Meteorological and environmental variables were normalized using Min-Max scaling and transformed into supervised learning sequences using a sliding-window approach. Baseline and dropout-enhanced models were evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The results show that dropout generally reduces forecasting error, although the magnitude of improvement varies according to architecture and dataset characteristics. The largest improvement was observed for RNN-drop on the Wellington dataset, with approximately 10.38% and 13.20% reductions in MAE and RMSE, respectively, while CNN-drop achieved the largest improvement on the Ohio dataset. These findings indicate that dropout can improve the generalization of established deep learning architectures, but its effectiveness is dependent on the interaction between model architecture and dataset characteristics. The study provides an empirical basis for selecting and evaluating dropout-regularized deep learning models for power consumption forecasting.