Improving Short-Term Heat Load Forecasts with Calendar and Holiday Data

Abstract

The heat load in district heating systems is affected by the weather and by human behavior, and special consumption patterns are observed around holidays. This study employs a top-down approach to heat load forecasting using meteorological data and new untraditional data types such as school holidays. Three different machine learning models are benchmarked for forecasting the aggregated heat load of the large district heating system of Aarhus, Denmark. The models are trained on six years of measured hourly heat load data and a blind year of test data is withheld until the final testing of the forecasting capabilities of the models. In this final test, weather forecasts from the Danish Meteorological Institute are used to measure the performance of the heat load forecasts under realistic operational conditions. We demonstrate models with forecasting performance that can match state-of-the-art commercial software and explore the benefit of including local holiday data to improve forecasting accuracy. The best forecasting performance is achieved with a support vector regression on weather, calendar, and holiday data, yielding a mean absolute percentage error of 6.4% on the 15–38 h horizon. On average, the forecasts could be improved slightly by including local holiday data. On holidays, this performance improvement was more significant.

Keywords

• district heating
• load forecasting
• machine learning
• weather data
• consumer behavior
• neural networks
• support vector machines