The transition to high shares of renewables leads to high demand of flexibility in the energy system. Due to increasing use of heat pumps and electric cars, the residential sector becomes more and more relevant in this respect. A precondition for reliably providing flexibility is the detailed knowledge of the current and future electricity demand of the respective household, which becomes possible with the comprehensive rollout of smart metering systems.
Since the load of individual households is highly volatile and stochastic, probabilistic instead of point forecast methods are considered. In this paper, different day-ahead probabilistic load forecasting methods for individual electricity consumption are developed. There are based on two different approaches; the first approach develops a point forecast which is subsequently transformed to an interval prediction. The second approach is a direct interval prediction. The point forecast methods are reference-based and use data from recent days and load profiles generated from historical data as input data to forecast the load of the next day. The direct interval forecast methods utilize only historical data. A naïve forecast considering the load of the last 24 hours is applied as benchmark.
These two approaches are compared and analysed with a case study of 74 households in Germany for one year.
The evaluations reveal that independently of the method, high interval widths are reached due to very volatile load patterns. Regarding the grade of fulfilment, the direct interval prediction methods show a more stable result. The variation of the grade of fulfilment between the different households differs around 6 % in comparison to more than 10 % for most point-based methods. Furthermore, the computation time for point-based methods is very high, while the direct interval prediction is quick and also needs less data. One advantage of the point-based method is the point prediction itself, when this is desired as additional output. All developed point prediction methods outperform the benchmark for all error evaluation.
As a next step, a comparison to alternative forecast methods based on artificial neural networks is intended. Furthermore, a detailed analysis about the performance of every single household is planned. This contributes to the research question whether specific load patterns and characteristics indicate forecasting performance.