Pattern Referencing for Hourly Temperature Forecasting: A Diurnal-Aware, Imputation-Free Model with an Expanding Reference Set

Abstract

Hourly near-surface air temperature over the Chiayi-Tainan coastal plain in southwestern Taiwan exhibits strong spatial variability and rapid transitions driven by sea-land breezes, radiative cooling, and diurnal forcing. This study revisits and extends an earlier pattern-referencing weighted k-nearest neighbor (WKNN) forecaster for hourly one-step-ahead prediction. The extensions include (i) using observed temperatures on the original physical scale (no mandatory normalization or standardization), (ii) incorporating diurnal-cycle predictors (sine and cosine of hour), and (iii) adopting an expanding reference set as new observations accumulate. Missing inputs are handled natively by computing similarity using only available components at each forecast time. The model is evaluated at 14 stations with year-long hourly forecasting in 2025 and is compared with persistence, an autoregressive integrated moving average (ARIMA) (3, 1, 3) benchmark, and an earlier baseline configuration using a fixed reference set, normalized inputs, and no diurnal predictors. Across stations, the proposed design reduces overall errors and improves tracking during rapid transitions. Case studies for a cold episode in January and the annual maximum-temperature episode in June show closer tracking of observed temperature evolution during rapidly changing conditions. An outage-prone period in August further demonstrates an operational advantage: Even under widespread missing observations across the station network, forecasts can still be issued whenever at least a subset of predictors is available. Overall, diurnal-aware pattern-referencing with direct inputs and an expanding reference set provides an accurate and deployable option for coastal station networks subject to intermittent missing observations. SIGNIFICANCE STATEMENT: Coastal weather in southwestern Taiwan can change quickly over short distances and hours, yet station networks often have gaps. We present a simple forecasting approach that predicts next-hour air temperature by finding past situations that most resemble the current one, while using only the measurements that are actually available at that moment. The method keeps temperatures in their original units, adds time-of-day information, and continuously updates its reference set as new observations arrive. Tested at 14 stations over 2025, it reduces errors overall and improves tracking during rapid cold and heat transitions. This improves real-time temperature guidance when data streams are unreliable.