Data-driven heating and cooling load predictions for non-residential buildings based on support vector machine regression and NARX Recurrent Neural Network: A comparative study on district scale
Section snippets
Introduction & state-of-the-art
The realisation of future sustainable energy systems is linked to the further development of district heating and cooling technologies, in order to increase energy efficiency, lowering costs and enable the implementation of smart energy systems. The 4th generation of district heating and cooling constitutes a key role to achieve these goals. An essential element in this context is the prediction of thermal loads [1,2].
Thermal load forecasting of buildings is mainly based on physical and
Dataset for the following case study
The non-residential district for demonstration is located in Germany and comprises about 200 building characterized by years of construction, that differ between 1918 and 2015. The data set comprises historic time series data from 01/01/2016 to 01/01/2018 with an hourly resolution. Table 1 shows the energy consumers within the district heating and cooling system, depending on building usage classes. In order to train and test, the data set has been split into two parts. The input vectors are
Results
In Table 5, the median of MAE and MSE values for heat load predictions are shown, depending on the forecast model and season. Table 6 illustrates the results for MAE and MSE, based on cooling load predictions.
Taking these performance indicators into consideration, the NARX RNN configurations show better performance, compared to the ε-SVM-R models for heating and cooling load predictions, which is exemplary illustrated by Fig. 8.
The latter shows a 40-h excerpt of predicted cooling loads in July.
Conclusion
In this district scale case study, the thermal load forecasting performance of ε-SVM-R models has been compared to the outcome of two NARX RNNs of different depth. Taking the results into consideration, the NARX RNNs outperformed the ε-SVM-R models based on the MAE and MSE. Although, the neural networks tend to overfit in case of user-driven cooling load time series due to their recursive abilities. The lowest accuracy concerning heating and cooling load predictions based ε-SVM-R models occur
Acknowledgements
The authors gratefully acknowledge the financial support by BMWi (German Federal Ministry for Economic Affairs and Energy) within the project Eneff:Stadt, Eneff:Campus: Living Roadmap (promotional reference: 03ET1352A). Particular thanks are due to the project partners for collecting and storing data used.
References (36)
- et al.
4th generation district heating (4GDH)
Energy
(2014) - et al.
Heat Roadmap Europe: combining district heating with heat savings to decarbonise the EU energy system
Energy Pol
(2014) Computer-aided building energy analysis techniques
Build Environ
(2001)- et al.
Comparison between detailed model simulation and artificial neural network for forecasting building energy consumption
Energy Build
(2008) Simple model for prediction of loads in district-heating systems
Appl Energy
(2002)- et al.
Multiple regression model for fast prediction of the heating energy demand
Energy Build
(2013) - et al.
Heat load prediction in district heating systems with adaptive neuro-fuzzy method
Renew Sustain Energy Rev
(2015) - et al.
A prediction model based on neural networks for the energy consumption of a bioclimatic building
Energy Build
(2014) - et al.
Heating, cooling, and electrical load forecasting for a large-scale district energy system
Energy
(2014) - et al.
Applying support vector machine to predict hourly cooling load in the building
Appl Energy
(2009)