This research introduces a novel series-parallel hybrid technique for air quality index (AQI) forecasting by combining an ARFIMA model with a parallel ensemble of LSTM, ConvLSTM, BiLSTM, and CNN architectures. A novel feature fusion method is suggested to improve time series forecasting. The hybrid model can capture linear and nonlinear patterns with statistical modeling and deep feature transformation, employing dimensionality reduction and additive/multiplicative hybrid strategies. The model generates both point and interval forecasts, with forecast error distributions modeled using t location scale and gaussian distributions to construct the prediction intervals. The method is assessed using four-point and four-interval accuracy measures against three hybrid baselines, five deep learning (DL) models, and thirteen machine learning (ML) models. Tests conducted using Delhi AQI data using different train-validation-test splits show that the suggested model performs noticeably better than other ML models at multiple significance levels, exhibiting greater prediction accuracy and resilience.