Abstract

Accurate energy consumption forecasting is crucial for managing and optimizing modern power systems, especially as renewable energy sources become more prevalent and grids grow more complex. This paper introduces a new hybrid model that combines Particle Swarm Optimization (PSO) with Long Short-Term Memory (LSTM) neural networks, significantly improving prediction accuracy. The main innovation is using PSO to dynamically tune LSTM hyperparameters such as layers, neurons, learning rate, and dropout rate, addressing the limitations of trial-and-error methods. The PSO-LSTM model is thoroughly tested with real data from the Roche Plate microgrid, showing superior results with an MAE of 59.885 Wh and an RMSE of 83.783 Wh, representing a 41.6% improvement over standalone LSTM models. By effectively modeling complex, nonlinear energy consumption patterns, this approach supports better decision-making and resource management for microgrids, especially in isolated areas with intermittent renewable energy. The findings highlight the model's potential to enhance smart grid management and facilitate the shift to sustainable energy systems.

Key words: Particle Swarm Optimization; LSTM; Energy Consumption Forecasting; Microgrid; Deep Learning