Energy consumption forecasting: A deep learning approach based on A-LSTM networks

In the face of global energy instability, driven by the geopolitical crisis between Ukraine and Russia, persistent inflation, the depletion of fossil fuels, and the climate emergency, understanding and anticipating energy consumption trends has become a strategic priority. This thesis tackles this issue by developing a forecasting model based on artificial neural networks, more precisely an A-LSTM architecture (Attention-based Long Short-Term Memory), to predict short-term global consumption of fossil and renewable energies. The study relies on historical data from covering the period 1965–2022 and including seven major world regions. Two datasets were constructed: one for fossil energy and the other for renewables. Each dataset contains seven input variables (regional consumption) and one output variable (global consumption). The data were analyzed and preprocessed using methods like normalization before being integrated into the model for training and prediction. The proposed A-LSTM model comprises an LSTM layer, a Dropout layer to limit overfitting, an Attention layer to weigh past time steps, and a Dense output layer. Training was conducted in Google Colab using TensorFlow/Keras with the Adam optimizer and EarlyStopping. The model demonstrated excellent generalization capabilities. It achieved respectively a Mean Absolute Error and Mean Absolute Percentage Error of 0,0002 and 1.77% for renewable energy and 0,0006 and 2.07% for fossil energy. This model has been then compared to some literature models like MLP, classical LSTM, ARIMA-ANN. The A-LSTM proposed consistently outperformed them, confirming the added value of attention mechanisms in enhancing time series forecasting. Forecasts to 2026 indicate continued growth in renewable energy consumption (average annual rate of 5.73%), while fossil energy remains stagnant (0.02%), suggesting a slow but steady global energy transition. this thesis demonstrates the high potential of the A-LSTM model in the field of energy forecasting, highlighting its accuracy, robustness, and adaptability; it offers an effective tool for policymakers and stakeholders in the energy sector.