A multi-step hybrid forecasting model based on optimized secondary decomposition and TCN-BiLSTM with residual learning

Abstract Accurate air quality forecasting is crucial for ecological environmental protection and public health early warning. To address the non-stationarity and noise inherent in AQI series, this study proposes a novel hybrid forecasting framework. This framework begins by processing the raw series through advanced decomposition methods namely a novel Chaotic Dynamic Neighborhood Grey Wolf Optimizer is proposed to adaptively optimize the parameters of Variational Mode Decomposition, enabling a preliminary and adaptive decomposition of raw series. Subsequently, a secondary decomposition using Singular Value Decomposition-Empirical Mode Decomposition is applied to the resulting components to suppress modal aliasing and extract purified features. Furthermore, the model incorporates multi-source auxiliary variables including meteorological factors. For the prediction phase, a hybrid model integrating a Temporal Convolutional Network and a Bidirectional Long Short-Term Memory network is constructed, with its hyperparameters automatically optimized using the Superb Fairy-wren Optimization Algorithm. The predictions for each component are finally aggregated into the output through a Residual Learning strategy. Experiments conducted in Beijing and Shanghai demonstrate that the proposed framework significantly outperforms benchmark models in short and long steps ahead forecasting tasks. Furthermore, a comprehensive ablation study involving eleven distinct model variants rigorously confirms that each constituent component, including the optimization algorithm, the secondary decomposition, and residual learning, contributes effectively and significantly to the overall forecasting accuracy.