Statistical and deep learning approaches to UV-B time series with multi-scale temporal patterns

초록

This paper presents a rigorous statistical investigation into the comparative performance of state-of-the-art deep learning architectures and classical stochastic models for forecasting time series characterized by multi-scale temporal patterns. We evaluate the Informer model, which utilizes a ProbSparse self-attention mechanism, against the traditional Seasonal Autoregressive Integrated Moving Average (SARIMA) framework, as well as several competing models (LSTM, BI-LSTM, Transformer and Facebook Prophet). Using daily maximum UV-B radiation data as a representative environmental proxy, we analyze how each model addresses seasonal structure and short-term temporal variations. Our results show that no single forecasting model is uniformly superior. Informer achieves the lowest overall MAE, whereas Transformer yields the lowest RMSE, with SARIMA and BI-LSTM also showing competitive performance. Relative model performance further varies across peak, transition, and trough periods, suggesting that the multi-scale temporal structure of UV-B radiation does not uniformly favor a single forecasting architecture. Pairwise Diebold–Mariano tests provide limited evidence of significant differences among the leading models, while residual diagnostics indicate remaining temporal dependence across all models. This study offers critical insights into the alignment between model inductive biases and the inherent stochastic properties of environmental data, suggesting that model selection for multi-scale environmental time series should account for the forecasting objective and seasonal regime rather than model complexity alone. © 2026 The Korean Statistical Society, and Korean International Statistical Society. All rights reserved.

키워드