short-term residential load forecasting: impact of calendar effects and forecast granularity - residential energy storage systems

The literature enriches the methods of "aggregate" load forecasting, which helps grid operators and retailers to optimize planning and scheduling.
Recently, the adoption of distributed generation and storage systems has increased, creating new demands for "classified" Load Forecasting of a single system
Customers are even device-level customers.
Obtaining high-resolution data from smart meters enables the research community to evaluate traditional load forecasting techniques and develop new forecasting strategies suitable for demand
Side classification load.
This paper examines how calendar effects, prediction granularity, and length of training sets affect the accuracy of the day
Advance load forecast for residential customers.
Square error (RMSE)
Normalized RMSE is used as the prediction error index.
Similar average RMSE results were obtained by regression tree, neural network and support vector regression, but statistical analysis showed that regression tree technology was significantly better.
The use of historical load distributions with daily and weekly seasonality, combined with weather data, makes the predictive power of explicit calendar effects very low.
In the settings studied here, it is shown that using a more coarse prediction granularity can reduce the prediction error.
It was also found that the one-year historical data was sufficient to develop a load forecasting model for residential customers, as further increases in the training data set had marginal benefits.