Heating, cooling, and electrical load forecasting for a large-scale district energy system

Load forecasting is critical for planning and optimizing operations for large energy systems on a dynamic basis. As system complexity increases, the task of developing accurate forecasting models from first principles becomes increasingly impractical. However, for large campuses with many buildings, the large sample size has a smoothing effect on the data so that aggregate trends can be predicted using empirical modeling techniques. The distinguishing features of this work are the large scale of the energy system (a college campus with approximately 70,000 students and employees) and the simultaneous forecasting of heating, cooling, and electrical loads. This work evaluates several different models and discusses each model's ability to accurately forecast hourly loads for a district energy system up to 24 h in advance using weather and time variables (month, hour, and day) as inputs. A NARX (Nonlinear Autoregressive Model with Exogenous Inputs) shows the best fit to data. This time series model uses a neural network with recursion so that measured loads can be used as a reference point for future load predictions. 95% confidence limits are used to quantify the uncertainty of the predictions and the model is validated with measured data and shown to be accurate for a 24 h prediction.