<![CDATA[Purpose: The purpose of this study is to develop hybrid deep learning model combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer mechanisms for short-term electricity consumption forecasting. The motivation arises from the limitations of existing studies, which are mostly based on single or dual hybrid models and have not fully captured complex temporal dependencies in electricity demand. To address this gap, the proposed CNN–LSTM–Transformer architecture is introduced to jointly model local patterns, sequential dependencies, and global temporal relationships, which have not been simultaneously explored in prior electricity forecasting studies. In addition, the study investigates the impact of multivariate inputs and feature engineering on forecasting performance. Methods: The proposed model is evaluated using the Tétouan City Electricity Consumption dataset, which includes load data from three zones and meteorological variables. Data preprocessing involves cleaning, Min–Max normalization, and sequence windowing for supervised learning. Model performance is assessed using MAE, RMSE, and MAPE. To ensure comprehensive evaluation, six experimental scenarios are designed, including univariate and multivariate settings, per-zone and combined-zone configurations, as well as feature selection-based scenarios, to analyze accuracy, robustness, and generalization capability. Result: The experimental results demonstrate that the proposed model achieves consistent forecasting performance across all scenarios. The best performance is obtained in the multivariate combined scenario (S4), with MAE values of 232–257 kWh, RMSE values of 385–433 kWh, and MAPE values ranging from 0.82% to 1.43% across all zones. The feature selection scenarios (S5 and S6) also show competitive performance, with MAE ranging from 234 to 324 kWh, RMSE from 393 to 590 kWh, and MAPE between 0.82% and 1.83%, indicating that engineered features can maintain prediction accuracy while reducing input complexity. In contrast, the univariate per-zone (S1) and multivariate per-zone (S3) scenarios produce higher errors, while the combined univariate scenario (S2) yields moderate improvements in certain zones. Overall, these findings confirm that integrating multivariate features with cross-zone data leads to the most accurate forecasting performance. Novelty: The novelty of this study lies in the proposed integration of CNN, LSTM, and Transformer architectures into a unified hybrid framework for electricity consumption forecasting, which has not been widely explored in the energy demand domain. The model effectively combines local feature extraction, sequential dependency learning, and global attention mechanisms, and demonstrates strong capability in handling both univariate and multivariate electricity consumption data, including additional meteorological variables. Practical Implications: This study shows that hybrid CNN-LSTM-Transformer can applied in electrical domain with multiple meteorological variables.]]>
