• • Irradiance dominates PV power output with a Pearson coefficient of r = 0.95, while humidity (r = −0.53) and temperature (r = 0.52) provide secondary predictive leverage; wind speed (r = 0.04) and wind direction (r = 0.03) fall below the 0.2 relevance threshold and are excluded, reducing input dimensionality and training convergence time without sacrificing accuracy.
• • Against a BiLSTM baseline, the TBiLSTM reduces MAE by 37.16%, MAPE by 55.59%, MSE by 52.11%, and RMSE by 33.75%, while raising R² by 4.03%—a margin that directly translates to lower imbalance penalties in day-ahead market bidding and reduced reserve procurement for grid operators.
• • Relative to CNN-BiLSTM, the hybrid architecture still achieves MAE reduction of 12.43%, MAPE reduction of 8.79%, MSE reduction of 20.98%, RMSE reduction of 11.13%, and R² improvement of 2.13%, demonstrating that multi-head self-attention captures cross-time-step dependencies that convolutional kernels cannot resolve under rapid cloud transients.
• • The largest single-model gain is against standalone CNN (MAE −47.63%, MAPE −68.80%, MSE −68.47%, RMSE −45.49%), confirming that pure convolutional feature extraction is inadequate for non-stationary PV power series and that recurrent bidirectional modeling remains essential for capturing diurnal ramp dynamics.