A Comparative Analysis of Time Series Forecasting Models and the Novel ELSTMD Approach for Renewable Energy Generation

Abstract

The increasing number of people and the huge industrial evolution have led to serious environmental consequences, which have made energy from Renewable Energy Sources (RESs) inevitable. However, these energy sources suffer from intermittency as they depend on natural sources, which change over time. Consequently, this research addresses the issue by proposing an ANN model for reliable renewable energy forecasting. The Encoding Long Short-Time Memory Decoding (ELSTMD) model consists of two encodings, one LSTM, and two decoding layers. Moreover, the introduced model and 21 time series forecasting models are evaluated, including statistical, Artificial Intelligence (AI), and hybrid approaches. While making predictions, the impact of 17 distinct features, categorized into five primary groups—electricity, weather, economic, and seasonality-related attributes, as well as natural disasters— has been analyzed. Thus, the models are trained as univariate time series with exogenous variables. To evaluate the model’s generalization, an additional dataset involving photovoltaic (PV) panels was used. The proposed model outperforms all competitors on both datasets, achieving an (R2, MAE, MSE, RMSE) of (0.56714, 0.06398, 0.00691, 0.08313) on the first dataset and (0.76574, 0.09344, 0.01417, 0.11903) on the second dataset. The most notable advantage of this research can be summarized as providing precise predictions for the electricity generated from RESs, which can help reduce the intermittency, randomness, and stability issues associated with these sources. It also plays a crucial role in optimizing electrical power systems’ planning and storage efficiency.