Training Size Matters: Impact of Training Data Size on Electrical Load Forecasting

Electrical load forecasting is an essential component of smart grid systems.The availability of historical data about the energy load of buildings can be restricted, for instance, because of the recently installed metering system.On the other hand, training the models with large datasets requires a lot of computation operations, thus a long execution time.The limited size of data and the desire for fast execution pose challenges for the training of forecasting models.In this paper, the six commonly used model architectures for short-term electrical load forecasting are compared based on their performance and execution time.In the experiments, different sizes of training datasets are considered.The results demonstrate that there is no universal model that achieves both the fastest execution time and the lowest error level.With a limited training dataset, autoregressive integrated moving average demonstrates the lowest errors.Random forest and support vector regression demonstrate higher error metrics with small training data sizes, however, faster execution times than autoregressive integrated moving average.If a large training dataset is available, the random forest demonstrates a good balance between performance and execution time in comparison to other models considered in the experiments. CCS Concepts General and reference