Dynamic characteristics of renewable energy generation and consumption using LSTM deep learning algorithm

In this study, we investigated the relationships between static and dynamic temporal variables that affect the generation and consumption of renewable energy through the analysis of nine datasets. These datasets comprise five univariate datasets depicting regional energy consumption and four multivariate datasets, namely, one from Tetouan City with weather data, an individual household (IndH) with smart meter data, wind, and solar generation data. We present a reproducible benchmarking workflow and best-practice evaluation protocol for short-term energy time-series forecasting across heterogeneous datasets. The pipeline combines exploratory statistical diagnostics and visualization with STL decomposition to characterize distributional properties and extract multi-scale seasonal components (hourly to quarterly), providing a consistent foundation for cross-dataset comparison. The results demonstrated different temporal characteristics for each renewable source; solar energy tracks with daily sunshine, peaking in the summer and declining in the winter, whereas wind energy showed more erratic patterns with peaks and declines during the winter months. Household energy consumption, in contrast to regional records, which show higher consumption on weekdays, was more reflective of daily activities and increases on weekends. Simultaneously, geographic climate also plays a massive role in this phenomenon, with seasonal extremes exacerbating the variation in energy use by increasing the heating or cooling demand during the winter or summer seasons, respectively. For robust and reliable predictive modeling, we recommend min–max normalization to preserve series shape during preprocessing, and using the sliding windows for time samples. An 80–90% training–testing split, complemented by an internal validation partition, then implementing five-fold cross-validation specifically adapted to time-series contexts using expanding windows with fixed validation sizes. The validation-enhanced LSTM forecasting model achieved an improvement of 0.6–2.2% compared to a non-validated configuration, justifying its robustness and making it a suitable candidate for deployment as the final model.