Combined Cooling, Heating, and Power Systems: Modelling Optimization, and Operation
5 Short-Term Load Forecasting and Post-Strategy Design for CCHP Systems
Download citation file:
- Ris (Zotero)
- Reference Manager
Having a designed operation strategy, another aspect that can restrict the CCHP system performance is the load profile. Almost all of the operation strategies are designed by assuming that accurate loads can be obtained in real time. However, this is not the case in reality. In practical applications, we cannot get full access to the heating, cooling and electrical loads in the following hour. The only information that can be used is the historical loads, current and historical dry-bulb, and dew-point temperature. Thus, the question arises naturally: Can we make use of the limited information to forecast the loads in the following hour? To forecast the load, the first step is to construct a forecasting model. In the literature, there exist many time series forecasting models, for example, AutoRegressive (AR) models, Moving Average (MA) models, AutoRegressive Moving Average (ARMA) models, AutoRegressive Integrated Moving Average (ARIMA) models, and AutoRegressive Moving Average with eXogenous inputs (ARMAX) models. The comparisons among different time series models can be found in [5, 6, 7]. Load forecasting can also be accomplished by adopting an artificial neural network (ANN) and Kalman filter. Other control approaches related to load forecasting and system identification can be found in [12, 13, 14, 15, 16], to name a few. Here, in this chapter, since the temperature information is available, the ARMAX model will be selected as the forecasting model; and the structure of two-stage least squares (TSLS) will be chosen to identify the model parameters. In order to improve the forecasting accuracy, instead of using the dry-bulb temperature only, the dew-point temperature will be considered as an instrument variable (IV). By doing so, the correlation between the dry-bulb temperature and the error term can be eliminated, and more weather information, such as the humidity, can be included inherently in the model. Thus, the estimated forecasting model can be more accurate than the one without using IV. The first stage of the parameter identification can be readily accomplished by ordinary least squares (OLS); since the second stage will involve a large amount of data, the two-stage recursive least squares (TSRLS), which has been proven to have a faster convergence rate and more accurate estimation, is adopted to reduce the space, computational, and time complexity. Combining OLS and TSRLS in one TSLS procedure is one of the objectives and contributions of this chapter.