Combining the strengths of different load modeling methods in short–term load forecasting of a distribution grid area with active demand

Power flows are becoming increasingly volatile in power distribution grids. Distributed power generation, electricity storage, electrical vehicles and active demand cause these variations. Dynamic management of constraints, power quality and balancing will be needed. The accurate forecasting of the power flows is a necessary enabler for it. It is important to accurately forecast the whole power balance including distributed generation, storage and loads. Much measurement data are now available which allows machine learning methods to become popular. A major challenge is that such black box methods are poor when significantly outside the situations included in the learning data. Especially in the presence of dynamic active demand, the black box models completely fail. Physically based model structures forecast those situations reasonably accurately, but require much domain expertise and development work. Modern load curve approaches fit well with grid state estimation and simulation, but like black box machine learning, the models require much more identification data than the models with physically based structures. Hybrid approaches aim at combining the strengths of different forecasting approaches and mitigating their individual weaknesses. This paper studies short–term load forecasting in a distribution area with about 9000 active consumers subject to both emergency and Time of Use load control. We integrate several modelling approaches to hybrid models to combine the strengths of the component models and avoid the weaknesses of the individual approaches. The component models include 1) models with physically based structures, 2) different machine learning methods, and 3) a similar day forecaster. We developed the methods and analysed their performance using field tests with load control actions and measurements from smart meters, distribution grid and weather. There are many model hybridisation methods. Here we mainly use sequential modelling of residuals. We begin with the physically based modelling and then model the residual using the other methods. The resulting forecast is the sum of the component methods. We also use physically based models to constraint the other forecasts to remain in a reasonable range and apply simple ensemble forecasting. The hybrid models were more accurate than the individual component models. The superiority is especially clear in exceptional situations and during dynamic load control actions. A further advantage is that the forecasting task for the other methods becomes easier when the models with physically based structures remove some fast and complex phenomena from the remaining forecasting task.