Atmospheric Ozone Prediction by Linear dendritic neuron model combining Seasonal-trend decomposition based on LOESS

Abstract Ozone in the atmosphere is not only closely related to people's lives and production, but can also endanger health at high concentrations. However, existing methods for predicting ozone levels are not accurate enough. Therefore, a prediction model composed of Seasonal-trend decomposition based on LOESS(STL) and improved dendritic neuron model(L-DNM) is proposed to predict atmospheric ozone content by dividing the time series into three parts: seasonal, trend and residual. Then, the trend part of the sequence is smoothed using the least squares method, and the most irregular residual is used as the input of the L-DNM. During the training process, the Back-propagation (BP) algorithm is used as the training algorithm to output the predicted part of the residual. The seasonal part with regularity and the other two processed parts are added together to obtain the prediction result. Through comparison, the STL-L-DNM is the state-of-the-art method in predicting ozone levels on five datasets considering four indicators including Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Nash-Sutcliffe Efficiency (NSE). This includes the most popular Transformer model and its variants in recent years. At the same time, The Taylor diagrams and regression diagrams show that the predicted value of STL-L-DNM is the most stable and the fluctuation is the least. The experimental results demonstrate the accuracy and stability of the proposed algorithm, which can play an important role in the prediction of ozone level.