Hourly temperature downscaling method based on clustering and linear transformation: Utilizing mean, maximum, and minimum temperatures

Abstract

The temperature profile under climate change is likely to differ from historical patterns. To understand how energy consumption and long-term load curves will be impacted by temperature variations, higher granularity temperature data is required. This study uses the AR6 statistically downscaled daily data provided by the TCCIP platform for Taiwan as a case study, proposing a method for downscaling daily temperature data to hourly data. The k-means algorithm clusters historical daily temperature profiles by month, using average, maximum, and minimum temperatures. Future temperature profiles are identified based on these characteristics, and a linear transformation is applied to align the downscaled hourly data. This method better captures the timing of maximum and minimum temperatures and the connection between daily profiles. Ultimately, 96% of the daily data, after downscaling, met the daily average, maximum, and minimum temperature values provided by TCCIP. Validation using data from January to June 2024 shows the method achieves an average absolute hourly error of 0.17–0.55 °C and a monthly average absolute hourly error of 0.2–0.4 °C, outperforming existing methods. The approach provides more comprehensive and accurate long-term temperature data, supporting studies on climate change impacts on energy demand, building design, and power system operations.