Transformer–LSTM Framework for Energy Consumption Prediction in Smart Energy Systems using Time-Series Deep Learning

Résumé

 Energy consumption forecasting is a key element in intelligent energy 
management systems and a smart grid environment. By accurately predicting 
electricity consumption, we can maximize energy utilization efficiency and 
reduce operating costs. This makes it important to predict energy consumption 
for developing a smart energy system. Traditional forecasting methods fail to 
adequately characterize the temporal dependencies and non-linearity within long
time-series data. The proposed framework in this paper aims to use a deep 
learning approach to establish a Transformer-LSTM model to perform short-term 
energy consumption forecasting. This Transformer-LSTM model combines the 
self-attention mechanism within Transformer networks and the sequence 
learning property within LSTMs to improve the forecasting performance and 
enhance temporal feature extraction. The proposed framework is evaluated using 
PJM Hourly Energy Consumption data and UCI Household Electric Power 
Consumption data and compared with existing models: LSTM, GRU, CNN
LSTM and Transformer models. Evaluating by MAE, RMSE, MAPE and R² 
Score respectively, the proposed hybrid framework demonstrated competitive 
forecasting performance across both datasets. The proposed model achieved the 
highest R² score and lowest RMSE on the UCI dataset, while maintaining 
comparable forecasting accuracy to recurrent and hybrid baselines on the PJM 
dataset. The proposed framework demonstrated competitive forecasting 
performance across both datasets while maintaining robust prediction 
accuracy on the PJM dataset. The approach can support intelligent decision
making within smart grids, smart buildings, cloud-based energy management 
systems, etc.