A comparison of local, cluster-specific and global Transformer models for forecasting electrical loads of individual buildings and substations

Forecasts of volatile electrical load are essential to operate smart buildings and the electrical grid. The number of measurement-equipped buildings and grid points increases due to the ongoing roll-out of measuring devices. Several strategies exist for training forecasting models for multiple time series from a given set of measuring points, including (1) local models, which are specialized for one time series; (2) global models, which generalize for all time series; and (3) cluster-specific models, which are applied to a cluster of time series with similar load patterns. We evaluate the three strategies for the recently developed time series Transformer forecasting model and compare them with linear regression (LR) and multilayer perceptron (MLP) models as benchmarks. We conduct experiments on two datasets, one with load data from 328 clients and one with load data from 152 high to medium voltage substations, thereby representing a variety of use cases. On both datasets, the Transformer gives the best results when it is used as a global model. An analysis of the global Transformer’s attention scores on different clusters of the data shows that the model adapts to different load patterns present in the clusters. The comparison with the baselines reveals that the global Transformer is the best model on the client dataset, but a local MLP is better on the substation dataset, indicating that the overall best-performing model depends on the characteristics of the dataset. We observe a tradeoff between forecast accuracy and energy usage for model training on the client-level dataset and outline future research on the sustainable usage of forecasting models for smart buildings.