District heating (DH) load forecasting for buildings and cities is essential for DH production planning and demand-side management. This study analyzes and compares the hourly DH load patterns for a city and five different types of buildings over an entire year. The various operating modes introduce nonlinear dependencies between the DH load and the outdoor temperature. We compare the prediction accuracies of different multiple linear regression (MLR) and artificial neural network (ANN) models. Without nonlinear dependencies, both ANN and MLR provide good, almost identical prediction accuracies. In the case of nonlinear dependencies, ANN is superior to MLR. However, the novel clustering method eliminates nonlinear dependencies and improves the accuracy of MLR on par with the ANN. ANN methods can automatically adapt to various nonlinearities. The advantage of combining MLR with the clustering method is that it is simpler than designing an ANN method, although manual work is required. In addition, MLR methods provide more insight into load patterns and how the load depends on various factors compared with ‘black-box’ ANN models. The developed methodology can be widely applied to building- and city-level load analyses and forecasting in different DH systems combined with or without domestic hot water consumption.