Forecasting of Heat Production in Combined Heat and Power Plants Using Generalized Additive Models

The paper presents a developed methodology of short-term forecasting for heat production in combined heat and power (CHP) plants using a big data-driven model. An accurate prediction of an hourly heat load in the day-ahead horizon allows a better planning and optimization of energy and heat production by cogeneration units. The method of training and testing the predictive model with the use of generalized additive model (GAM) was developed and presented. The weather data as an input variables of the model were discussed to show the impact of weather conditions on the quality of predicted heat load. The new approach focuses on an application of the moving window with the learning data set from the last several days in order to adaptively train the model. The influence of the training window size on the accuracy of forecasts was evaluated. Different versions of the model, depending on the set of input variables and GAM parameters were compared. The results presented in the paper were obtained using a data coming from the real district heating system of a European city. The accuracy of the methods during the different periods of heating season was performed by comparing heat demand forecasts with actual values, coming from a measuring system located in the case study CHP plant. As a result, a model with an averaged percentage error for the analyzed period (November-March) of less than 7% was obtained.