Performance Benchmarking of a Novel CNN–BiLSTM–Attention Hybrid Model for Short-Term Spatiotemporal Wind Speed Forecasting Using Indian SCADA Data

Abstract

Short-term wind speed forecasting is fundamental to the reliable integration of variable renewable energy into modern power systems. Although numerous machine learning and deep learning models have been proposed, comprehensive benchmarking of hybrid attention-based architectures against both classical and modern baselines on real Indian SCADA data accompanied by thorough residual diagnostics and statistical validation remains scarce. This paper reports a rigorous performance benchmarking of a CNN–BiLSTM–Attention hybrid model against five established baselines: Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), Support Vector Machines (SVM), hybrid ARIMA–ANN, and standalone CNN–BiLSTM. Using a cleaned 8,760-hourly SCADA dataset from an operational onshore Indian wind turbine, the proposed hybrid achieved MAE = 0.824 m/s, RMSE = 1.146 m/s, MAPE = 13.7%, and R² = 0.924 on the held-out test set, corresponding to RMSE reductions of 22.4%, 18.7%, 23.8%, 24.0%, and 11.2% versus ANN, LSTM, SVM, ARIMA–ANN, and CNN–BiLSTM respectively. Residual diagnostics Q-Q plots, Ljung–Box, Shapiro–Wilk, and Breusch–Pagan tests confirm white-noise behaviour with no systematic autocorrelation or heteroskedasticity. The Diebold–Mariano test verifies that observed improvements are statistically significant (p < 0.001 against all baselines). Wind-regime stratification further reveals that the hybrid's advantages are most pronounced in high-wind and transitional conditions. The findings establish the CNN–BiLSTM–Attention architecture as the most accurate and statistically robust candidate for operational deployment in Indian wind corridors.