Hybrid Deep Learning Models With Attention Mechanisms For Short-Term Wind Speed Forecasting: A Comprehensive Review And Benchmarking Of CNN–Bilstm–Attention On Real SCADA Data

Abstract

Wind speed forecasting is a critical enabler of reliable renewable energy integration, particularly in countries such as India where installed wind capacity exceeds 40 GW and continues to grow under ambitious 2030 renewable targets. While hybrid deep learning architectures combining convolutional, recurrent, and attention components have shown promise on synthetic and benchmark data, real-world validation on operational SCADA data remains essential for translating such models into deployable forecasting tools. This paper presents a comprehensive review of hybrid deep learning models with attention mechanisms for short-term wind speed forecasting and reports a full benchmarking study of the CNN–BiLSTM–Attention architecture on real SCADA-recorded data collected from an onshore Indian wind turbine over the calendar year 2018. The dataset comprises 52,592 ten-minute observations resampled to 8,760 hourly samples, with chronological 80/20 train-test splitting. Six forecasting models ARIMA, SVR, Random Forest, XG Boost, LSTM, and CNN–BiLSTM–Attention were compared using MAE, RMSE, MAPE, and R². The hybrid model achieved the best real-SCADA performance with MAE = 0.824 m/s, RMSE = 1.146 m/s, MAPE = 13.7%, and R² = 0.924, outperforming all baselines. Residual diagnostics using the Ljung–Box, Shapiro–Wilk, and Breusch–Pagan tests confirmed white-noise error behaviour, while the Diebold–Mariano test verified statistical significance of the observed gains. These results establish operational viability of the hybrid architecture for Indian wind corridors and provide empirical benchmarks for grid integration and regulatory-compliance applications.