Predicting and analysing flight delays is essential for successful air traffic management and control. We propose a novel parallel-series model and novel adaptive bidirectional extreme learning machine (AB-ELM) method for prediction and feature analysis to better understand the causes of flight delays as stated by the International Air Transport Association (IATA). The IATA-coded flight delays are rarely examined in the existing studies. The IATA-coded flight delay subcategories decision boundaries are improved by the proposed parallel-series model. In application areas, where multiclass-multilabel classification may produce erroneous performance, the parallel-series model can be regarded as an alternate strategy. To improve network generalization performance, the proposed AB-ELM optimizes the covariance objective function by altering the learning rate adaptively during gradient ascent as opposed to gradient descent. The historical data from one of Hong Kong's international airlines, which contains information about the airport, flight, aircraft, weather, and IATA flight delay subcategories is considered a case study. Using fourteen different sampling approaches, the influence of imbalanced and noisy data was reduced. The results showed that employing proper sampling approaches in conjunction with the parallel-series model and AB-ELM method is effective for uncovering hidden patterns in the complicated IATA-coded flight delay subcategories system. When compared to other data-driven approaches, AB-ELM attained a high accuracy of 80.66 percent. This study enables airlines to develop adequate contingency measures in advance based on potential flight delay reasons and duration.