Short-Term Heat Load Forecasting in District Heating Systems : A Comparative Study of Various Forecasting Methods

Abstract: Short term heat load forecasts are vital for optimal production planning and commitment of generation units. The generation utility also bares a balance responsibility toward the electricity market as a result of CHP generation. Sub-optimal load forecasts can lead to high costs relating to unit commitment, fuel usage and balancing costs. This thesis presents the empirical comparison of various models for 24h heat load forecasting. Five methods were investigated including four supervised machine learning algorithms; neural networks, support vector machines, random forests and boosted decision trees and one auto-regressive time series model; ARIMAX. The models were developed, and evaluated using cross validation with one year of hourly heat load data from a local district heating system and corresponding meteorological data from the same time period. The thesis also investigates the impact of feature selection on the predictive power and generalization ability of the models. The results indicate a significant difference in forecast accuracy between the methods with neural networks and ARIMAX showing the best and similar performance followed by the support vector machine, boosted decision trees and random forest.