Forecasting Heat Demand with Complex Seasonal Pattern Using Sample Weighted SVM

Short-term forecasting of heat demand is crucial for controlling district heating networks and integrated electricity and heat supply systems. The forecast specifies an estimate of the energy required in the coming hours which enables the controller to proactively manage the storage units and schedule the heat generation. Consequently, improving the accuracy of heat demand forecasting can lead to reduced operational cost and increased reliability of the energy supply. This paper presents the development of a sample weighted Support Vector Machine (SVM) to improve the accuracy of heating demand forecasting. As the dynamics of heat demand time series change over time, recurrence plot analysis is first used to investigate any seasonal behavior and its relationship to ambient temperature. Then, to capture this seasonal behavior, a membership-function-based method is presented to generate the weight of each sample in learning a SVM model. This method is evaluated using a dataset with half hourly resolution from an industrial case study in the UK. Compared to conventional forecasting methods, the proposed approach shows significantly better accuracy in 24 hours ahead forecasting of heat demand.