An Efficient Joint Evolutionary Algorithm-based Neural Network Model for Air Quality Prediction

The Air Quality Index (AQI) prediction is an important tool for assessing and managing air pollution, with broad application value. In recent years, deep learning methods have compensated for the deficiencies in prediction accuracy of traditional AQI prediction methods due to their advantages in handling nonlinear data and feature learning. However, existing deep learning-based AQI prediction methods have not adequately considered temporal logic during the training process and lack differentiated handling of different types of features. Furthermore, the hyperparameter tuning process of network models is complex and inefficient. To address these issues, this paper presents a joint optimization network model (JONM) based on adaptive weights to achieve more accurate air quality prediction. The main innovations include four aspects: a) a novel spliced multi-step prediction method is introduced, which segments the input sequence for prediction, ensuring non-iterative accumulation of errors while fully considering the temporal relationships of the predicted sequence, thereby resolving the error iteration problem in traditional multi-step predictions, b) the input feature data is processed differentially by classifying the input features into primary features (AQI) and secondary features, allowing different features to flow through neural network layers of varying depths. This ensures that while the model learns secondary information to assist in predictions, AQI prediction remains the primary task, avoiding interference between primary and secondary tasks, c) a joint evolutionary algorithm is constructed using eight evolutionary algorithms with different mutation strategies, guiding the model's prediction results by weighted summation of the hyperparameter groups optimized by each evolutionary algorithm, d) particle swarm optimization is utilized to optimize the joint weights of the prediction results from the eight evolutionary algorithms, thereby enhancing computational efficiency. Experiments and evaluations on real AQI datasets indicate that the proposed JONM outperforms the benchmark model, validating the effectiveness of the joint strategy and the accuracy of the prediction results.