Implementation and Validation of an Air Quality Prediction Model Based on Federated Learning

Abstract

Ground ozone and other air pollutants pose a great threat to the health. Because the data of different monitoring stations are scattered, it is very complicated to accurately predict the ozone level. Federated Learning (FL) allows everyone to train models together without exchanging raw data, thus protecting privacy. This paper uses FL to predict air quality, and the key point is to establish a prediction model, not to change the algorithm. This paper used UCI air quality data set (with 8,762 records, each with four attributes) to simulate the situation of five users, and the data were separated by independent identically distributed (IID) and non-independent identically distributed (Non-IID). This paper trained 50 communication rounds with a multi-layer perceptron (MLP) and Federal Average (FedAvg) method. The model of centralized training and local training is also used to compare. The experimental results show that the average absolute error (MAE) of FedAvg is 48.9 under IID data and 58.3 under Non-IID data, which is much better than the local training models (MAE 78.5 and 85.2) and close to the effect of centralized training (MAE 42.3 and 51.6). This method of federated learning adapts well to different data. This study confirms that FL is a feasible method for distributed air quality prediction.