This study applies statistical methods and deep learning techniques to forecast the daily average PM2.5 concentration in northern Thailand, where the concentration is usually high and exceeds the safe level. The data used in the analysis are collected from January 2018 to December 2020 from 16 air monitoring stations. The statistical methods used are Holt-Winters exponential smoothing (ETS), autoregressive integrated moving average (ARIMA), and dynamic linear model (DLM). The deep learning techniques considered in this study are the recurrent neural network (RNN) and long-short term memory (LSTM). To compare the predictive performance of both methods, we use the root mean square error (RMSE). The result indicates that statistical methods, especially ARIMA, perform better than the deep learning techniques in most stations. Moreover, LSTM tends to provide higher accuracy than the RNN, especially with more number of nodes.

This study proposes a new five-parameter distribution called the gamma-exponentiated Weibull–Poisson (GEWP) distribution. As an extension of the exponentiated Weibull–Poisson family, the GEWP distribution offers a more flexible tool for analyzing a wider variety of data due to its theoretically and practically advantageous properties. It encompasses established distributions like the exponential, Weibull, and exponentiated Weibull. The development of the GEWP distribution proposed in this paper is obtained by combining the gamma–exponentiated Weibull (GEW) and the exponentiated Weibull–Poisson (EWP) distributions. Therefore, it serves as an extension of both the GEW and EWP distributions. This makes the GEWP a viable alternative for describing the variability of occurrences, enabling analysis in situations where GEW and EWP may be limited. This paper analyzes the probability distribution functions and provides the survival and hazard rate functions, the sub-models, the moments, the quantiles, and the maximum likelihood estimation of the GEWP distribution. Then, the numerical experiments for the parameter estimation of GEWP distribution for some finite sample sizes are presented. Finally, the comparative study of GEWP distribution and its sub-models is investigated via the goodness of fit test with real datasets to illustrate its potentiality.

Rainfall regionalization plays an essential role in identifying homogeneous rainfall patterns and supporting hydrological and climate analyses. In Thailand, the regionalization adopted by the Thai Meteorological Department (TMD) is primarily based on monsoon wind systems and broad geographic boundaries, which may not adequately represent sub-regional variability in rainfall behavior. This study proposes a data-driven framework to identify homogeneous rainfall regions using monthly rainfall observations from 67 stations across Thailand over the period 1983–2018. Two representations are examined: a standardized representation and a principal component analysis (PCA)-based standardized representation. K-means clustering is applied to both representations, and the resulting clusters are then evaluated using L-moment homogeneity testing, principal component visualization, and statistical and spatial validity indices. The PCA-standardized dataset produces clusters with improved separation and stronger homogeneity relative to the standardized dataset, and the resulting rainfall regions are further interpreted in terms of their physical rainfall characteristics. The practical relevance of the identified regions is further demonstrated through leave-one-out cross-validation comparing inverse distance weighting (IDW), K-nearest-neighbor IDW, and cluster-based IDW approaches using the proposed regions, existing data-driven regions, and TMD regions. The proposed cluster-based IDW approach achieves interpolation error reductions of approximately 9.18–11.55% compared with conventional IDW- and TMD-based alternatives, while providing performance comparable to that obtained using other recent data-driven regional classifications.