A Pattern-Referencing Model for Hourly Temperature Forecasting in Coastal Regions

Abstract

This study proposes a pattern-referencing model for hourly temperature forecasting in coastal regions, specifically designed for scenarios with missing data. The Chiayi-Tainan coastal plain in Taiwan exhibits pronounced spatiotemporal temperature variations driven by sea-land breezes, topography, and solar radiation, impacting real-time decision-making in industries such as aquaculture, agriculture, and tourism. The proposed model directly utilizes all available input data without requiring prior imputation or specialized pretraining. In a multistation study involving 14 weather stations, the model employs a weighted K-nearest neighbors (WKNN) approach, using a masked Euclidean distance and the Dudani weighting scheme. The optimal configuration (look-back length = 1, number of neighbors = 18) achieved mean absolute errors of 0.35 degrees C-0.59 degrees C and root-mean-square errors of 0.45 degrees C-0.86 degrees C across diverse weather scenarios, outperforming both persistence forecasts and an autoregressive integrated moving average (ARIMA) model. The model performs best under low-temperature conditions but shows a slight tendency to underestimate at high temperatures; nighttime forecasts are the most stable, while daytime errors are larger. Even with missing station data, the model maintains its predictive capability, offering decision-makers more reliable hourly forecasts in resource-limited networks with unstable data availability, and enabling policymakers to build early-warning systems that help coastal communities and industries respond to extreme temperature events.