Conference Year



optimization, load forecasting, prediction, model predictive control, regression models, day-typing


The reduction of energy consumption, use of renewable energy, and preservation of natural resources are becoming increasingly important. Several applications in the energy efficiency field aim at minimizing energy consumption and/or cost. To achieve this, these applications employ optimization techniques that require future prediction of the performance and various loads of a facility, campus, building, or an energy plant, such as hot water, cold water, and/or electric load. The prediction horizon may be as short as few hours to ten days into the future, depending on the application at hand. Furthermore, for the purpose of minimizing electricity cost, it is as necessary to know, as accurately as possible, what the electricity rates are over a given horizon. Therefore, a method for predicting hot water, cold water, and electric loads and electricity rates over a given horizon into the future has been developed (Load will be used to refer to hot water, cold water, and electric loads and electricity rates without loss of generality). The method developed takes into consideration the several factors contributing to the load value. These factors include time of day, day of week, schedules (in-session or out-of-session for a university campus for example), and weather (temperature and humidity). The load predicted consists of a deterministic term and a stochastic term. The deterministic term is calculated using linear regression models, whose coefficients are determined offline. These models rely on the typical load value for a given time of day and day-type (days with similar load profiles) and weather forecast. The latter is obtained from the National Oceanic and Atmospheric Administration (NOAA) through their National Digital Forecast Database (NDFD) service. The stochastic term is determined using an Auto-Regressive (AR) model, whose coefficients are determined offline. The AR model calculates future prediction errors based on the current prediction error. The stochastic element of the predicted load gives the method developed its adaptive property, and thus increases the accuracy of the prediction by updating the forecast using current measurements of the load. Historical weather and load data are used for determining the coefficients of the regression models and the AR model offline. For a given set of training data, the method developed generates a set of regression models for each day-type. Day-types are determined by a day-typing algorithm which specifies days with similar load profiles based on cluster analysis techniques. Outside air enthalpy and a typical load profile constitute the predictors variables in each set of regression models. Each day-type is characterized by a different typical load profile which is generated using an optimal data fitting technique. The AR model coefficients are determined using the residuals obtained from different sets of regression models. Given the determined models, the current load measurement, and weather forecast, the future load values are calculated by selecting the appropriate regression model and summing the deterministic and stochastic terms.Â