Flight delays are a chronic problem in the modern air transport system, affecting not only the operational efficiency and service quality of airlines, but also the transport efficiency of millions of passengers and goods. Due to the irregularity and uncertainty of the time series of flight data, the current methods lack further exploration and mining of time series characteristics. This paper proposes a novel hybrid bidirectional temporal convolutional (BiTCN) network model: CPO-BiTCN-BiGRU-Attention. The model is based on a dynamic sliding window which divides the airport delay time series into time windows of different sizes on a daily, weekly and monthly basis to effectively capture the delay patterns at different time scales. The processed time series are subsequently input into the BiTCN-BiGRU structure, where the Crowned Porcupine Optimization (CPO) algorithm is employed to optimise the network structure parameters. This bidirectional structure enables the model to learn both past and future temporal dynamics and relationships within the sequence, enhancing its capacity to identify periodic patterns and improving responsiveness to temporal fluctuations. Finally, the multi-head attention mechanism assigns differential importance to hidden states from distinct temporal perspectives. Experiments were conducted using a flight dataset from Atlanta International Airport (ATL) in the year 2023. The top ten airports with the highest flight volume at ATL were selected. Compared with the existing benchmark method, the hybrid BiTCN model achieves approximately 15% and 3% reductions in MAE and RMSE, respectively, relative to the conventional convolutional neural network model.