Machine Learning Approaches for Renewable Energy Forecasting in Smart Grids: A Hybrid Deep Learning Framework with Attention Mechanisms

Abstract

The accelerating integration of renewable energy sources (RES) into modern power systems has created an urgent need for accurate short-term forecasting techniques capable of capturing the intermittent and nonstationary nature of solar irradiance and wind generation. Traditional statistical methods such as ARIMA and exponential smoothing fail to model the complex nonlinear dependencies between meteorological drivers and power output, while standalone deep learning architectures often struggle with long-range temporal correlations and feature heterogeneity. This article proposes a hybrid deep learning framework that combines a one-dimensional convolutional neural network (1D-CNN) feature extractor, a bidirectional long short-term memory (BiLSTM) sequence encoder, and a multi-head self-attention mechanism for short-term renewable energy forecasting in smart grid environments. The CNN block (64 filters, kernel size 3) extracts local spatiotemporal patterns from weather and lag features; a BiLSTM with 128 hidden units captures bidirectional sequence dependencies; and an 8-head self-attention layer dynamically reweights the most informative time steps before a dense regression head produces multi-horizon forecasts. The framework was evaluated on three real-world datasets: a 10 MW solar photovoltaic (PV) plant in Rajasthan, India (two years, 15-minute resolution); a 50 MW wind farm in Sindh, Pakistan (two years); and a 25 MW solar plant in Luzon, Philippines (one year). Against seven baselines, including persistence, ARIMA, support vector regression, random forest, vanilla LSTM, vanilla GRU, and CNN-LSTM, the proposed model achieved a normalized root mean square error (NRMSE) of 0.064 on the Indian solar dataset, compared with 0.121 for persistence and 0.094 for vanilla LSTM, representing a relative improvement of 47.1% and 31.9% respectively. Ablation studies confirm that both the attention mechanism (NRMSE rising to 0.078 without it) and the CNN block (0.082 without it) contribute materially to performance. The model achieves an inference latency of 12 ms per sample, supporting deployment in real-time energy management systems.