PM2.5 Prediction Based on LSTM Weighted by K-Nearest Neighbor Algorithm

Accurate prediction of PM2.5 concentration is essential for public health and environmental protection, and specifically crucial for the management of the availability of sufficient health personnel during adverse health episodes. However, its nonlinearity, variability, and complexity make this task challenging. This study proposes a long short-term memory (LSTM) weighted by K-nearest neighbor (KNN) algorithm (namely Weighted KNN-LSTM Model) that can effectively predict the PM2.5 concentration time series. Firstly, the K-nearest neighbors of each time point are sought based on the Euclidean distance within the data time range. Given that neighboring observations typically exert a more pronounced influence than distant ones in spatial processes, weights are accordingly assigned to these neighbors to quantitatively reflect their relative importance in the analysis. Subsequently, after the initial data is processed by the weighted KNN algorithm, it is reorganized and transformed into a reconstructed dataset with a size K times that of the original data. The data used for model training and the data used for evaluating the model’s prediction performance are completely independent, and the test dataset is never involved in the model training process to ensure the authenticity and reliability of the prediction performance evaluation. Then, the LSTM neural network model is trained on this new dataset to enhance its generalization ability. The experimental results show that the weighted KNN-LSTM model exhibits excellent predictive performance in predicting PM2.5 concentration. It is important to note that the dataset used to evaluate the model’s performance was strictly independent from the data used to train the model. This separation ensures that the reported accuracy reflects true predictive capability rather than mere fitting quality. The model provides a technical reference for hourly PM2.5 concentration prediction in Nanchang City, and the prediction results can be used as an auxiliary reference for regional air quality monitoring; the application of the model in heavy pollution warnings needs to be further optimized and verified by combining multi-source data such as meteorology, which provide reliable data support for the formulation of dynamic emission reduction policies.