Analysis of heating degree-days (HDD) data using machine learning and conventional time series methods

Heating degree-days (HDD) is a unit to measure the duration of cold weather during a 24-h period. Accurate HDD forecasting is crucial in terms of determining the energy needed for the heating of buildings. This work aims to introduce a model for one month-ahead forecasting of HDD time series data and their short-term future projection. Seasonal autoregressive integrated moving-average (SARIMA) and long short-term memory (LSTM) neural network models are utilized for this aim. Monthly HDD time series data from 2007 to 2021 in different climatic regions of Turkey are supplied by the Turkish State Meteorological Service (TSMS). Seasonally non-stationary HDD data for both models are converted to the position of stationary time series. Akaike information criterion (AIC) and Bayesian information criterion (BIC) values are utilized to optimize the models. The evaluation metrics of root-mean-square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) are used to assess the forecast accuracy of the models. As a result, forecasts with approximate 95% confidence limits are made for the 60-month period (January 1, 2022 to December 31, 2026) from the proposed models. The results show that the LSTM and SARIMA models can provide future forecastings for HDD time series data with satisfactory precision.