Elsevier

Energy

Volume 87, 1 July 2015, Pages 343-351
Energy

Forecasting of consumers heat load in district heating systems using the support vector machine with a discrete wavelet transform algorithm

https://doi.org/10.1016/j.energy.2015.04.109Get rights and content

Highlights

  • District heating systems application for increasing in fuel efficiency.

  • Control and prediction of heat load for individual consumers in district heating systems.

  • Support vector machine application for district heating system control algorithm.

Abstract

District heating systems are important utility systems. If these systems are properly managed, they can ensure economic and environmentally friendly provision of heat to connected customers. Potentials for further improvement of district heating systems' operation lie in the improvement of current control strategies. One of the options is the introduction of model predictive control. Multi-step ahead predictive models of consumers’ heat load are a starting point for creating a successful model predictive strategy. For the purpose of this article, short-term multi-step ahead predictive models of heat load of consumers connected to a district heating system were created. The models were developed using the novel method based on SVM (Support Vector Machines) coupled with a discrete wavelet transform. Nine different SVM-WAVELET predictive models for a time horizon from 1 to 24 h ahead were developed. Estimation and prediction results of the SVM-WAVELET models were compared with GP (genetic programming) and ANN (artificial neural network) models. The experimental results show that an improvement in predictive accuracy and capability of generalization can be achieved by the SVM-WAVELET approach in comparison with GP and ANN.

Introduction

District heating is a utility energy service based on the provision of heat to remote customers (connected to the system via heating substations) from available heat sources [1]. It enables utilization of waste or low-cost heat, which is the main precondition for DHS (district heating system) competitiveness, when compared to onsite, individual, boilers. Another precondition is high heat density. Although some previous studies showed that sparse DHS (systems with low heat density) can be economic [2], [3], [4], their success correlates highly with energy taxation which is country specific, even in the EU. However, according to [5], [6], 73% of EU 27 residents (502 million) lived in urban areas, which indicates high heat density and excellent prospects for further growth of DHSs in Europe. In Ref. [7] the potential for district heating in the EU from now to 2050 was identified, based on extensive and detailed mapping of the EU heat demand and various supply options. The results in Ref. [7] indicated that with district heating, the EU energy system will be able to achieve the same reductions in primary energy supply and carbon dioxide emissions as the existing proposed alternatives. Even in China coal boilers are still prevalent [8]. Analysis in Ref. [9] showed that a substantial reduction in fuel demands and CO2 emissions, as well as cost can be achieved by converting the heat from individual boilers based on oil, natural gas, or biomass to district heating. Such a conclusion seems to be valid in the present energy systems as well as in a future scenario aimed towards a 100 per cent renewable energy supply in 2050, even if the space heating demand is reduced to as much as 25 per cent of the present demand. District heating plants are able to resolve the conflict of must-run generation of base load power plants for control power provision; therefore, they can enhance the integration of fluctuating renewable energy sources [10].

The basic assumption in Ref. [11] was that DHS has an important role to play in future sustainable energy systems, including 100 percent renewable energy systems, but the present generation of DHS technologies will have to be developed further into a new generation in order to play such a role. One possible way for further increase of DHS competitiveness lies in the improvement of the currently used control strategy. The study in Ref. [11] suggested that intelligent control systems could be used to improve the efficiency of the DHS. Research [12] presented a scientific tool for the optimization of design and operation for complex polygeneration plants serving a number of buildings with heat, cooling, and electricity. The prevailing control strategy is based on weather compensation control, where the primary supply temperature (temperature of water from the heat source) is determined from the so-called “sliding diagram” [13]. This diagram provides the functional dependence between the momentary outdoor temperature and primary supply temperature of water pumped in DH network. Correlation between these two variables is almost linear, with two or three knee points, and often corrected by operators after years of usage.

However, district heating systems are complex and dynamic systems with high inertia and marked heterogeneity of users, and this precondition of static correlation between the outdoor temperature and heat pumped in the network does not hold. As a consequence, primary return temperature is frequently higher than needed, which is especially evident during the periods of moderately cold weather. Therefore, in order to reach the higher efficiency in DHS operation, the control strategy should be directed towards lowering of the primary return temperatures. Lowering of the primary return temperature in DHS has the following effects on the overall operation: increased electrical output from CHP-plants (usually used as heat sources), increased heat recovery from industrial excess heat and geothermal heat, lowered distribution losses and increased coefficient of performance if heat pumps are used in heat generation [14]. Economic benefit of reduced primary return temperature was estimated at 0.05 to 0.5 €/MWh for 1 °C of reduced temperature [15].

In order to get the lowered primary return temperatures, the produced and delivered heat from the heat supply units must closely correspond to the heat demanded by consumers, increased for appropriate distribution losses. Taking into account that changes in supply temperature in the heat source will be “sensible” at the peripheral parts of the network after considerable time has passed, due to transport delay (water temperature changes “travel” at the speed of water in the network) in the distribution network, the required heat to be pumped in the network should be known in advance (several hours ahead), in order to avoid excessive heat input in the network with a very important precondition that consumers get the required heat. Therefore, predictive heat load models of all, or at least the most influential, consumers in the system are indispensable as inputs for advanced model predictive control. Predictive models should provide several-hours-ahead forecasts of required heat, where the prediction horizon can be defined according to the endmost consumer in the network.

In this article, we introduce the heat load prediction models of consumers for a different prediction horizon using the data acquired from one heating substation in DHS Novi Sad, Serbia. The proposed models are developed using a soft computing approach, namely SVM (Support Vector Machine) with a discrete wavelet transform algorithm, which are presented and compared.

SVMs (Support vector machines) are a type of soft computing technique that has gained importance in different engineering disciplines [16], [17]. The prediction accuracy of an SVM model heavily relies on proper determination of model parameters. Although organized strategies for selecting parameters are important, model parameter alignment also needs to be made. In the past, different researchers applied various conventional optimization algorithms for selecting these parameters, but with limited success [18], [19], [20], [67]. These include the grid search algorithm [21] and the gradient decent algorithm [22], [23], [68]. Computational complexity seemed to be the major drawback of the grid search algorithm, which restricted its applicability to simple cases. Gradient search algorithm, on the other hand, is usually prone to local minima. In most optimization problems, multiple local solutions do exist, but evolutionary algorithms seem to be the best approach, because they are capable of providing a global solution to such optimization problems.

In this study, a short-term predictive model of consumers heat load was developed using SVMs with the discrete wavelet transform algorithm (SVM-WAVELET). Wavelet analysis was used to decompose the time series of data into its various components, after which the decomposed components can be used as inputs for the SVM model. The results indicate that the proposed model can adequately predict the heat load for different prediction horizons. SVM-WAVELET results were also compared with GP (genetic programing) and ANN (artificial neural network) results.

The organization of the remaining part of this paper is as follows: Section 2 explains the heat load data and prediction models of consumers for different prediction horizons. Section 3 presents the description of the SVM-WAVELET method and GP and ANN used as benchmark methods. The comparative results and discussion are presented in Section 4. Finally, the conclusions are presented in Section 4.

Section snippets

System and data description

Collection of data used for subsequent analysis and building of the prediction models was conducted in the district heating system in Novi Sad, Serbia. The district heating system of Novi Sad is the second biggest DHS in Serbia with five heat supply units (Table 1) and one main manifold used for transferring the heat from the CHP unit “Novi Sad” towards the three supply units within the city and back (Fig. 1).

Base load for consumers connected to heat supply units “East”, “North”, and “South” is

Results and discussion

Predictive models of heat load were developed for: 1, 2, 3, 4, 5, 8, 10, 12, and 24 h ahead. Five models, for the prediction horizon from 1 to 5 h ahead, are intended for control purposes, while the models for 8, 10, 12, and 24 h ahead are intended for planning. When developing multi-step ahead prediction models, two approaches are possible:

  • Iterated prediction approach and

  • Direct prediction approach.

In the iterated prediction scheme, one-step ahead prediction is used for building the subsequent

Conclusion

Predictive models for consumers’ heat load in DHS were developed using the SVM-WAVELET method. The proposed SVM-WAVELET model was obtained by combining two methods, i.e. SVM and WAVELET transform. SVM performs structural minimization whereas other traditional, soft-computing, techniques focus on the process of minimizing errors, which is far less efficient. This proposed combination of methodologies is unique, and it enhanced the performance of the proposed model compared with models that were

Acknowledgment

This paper is financially supported by the Malaysian Ministry of Education under the University of Malaya High Impact Research Grant – UM.C/HIR/MOHE/FCSIT/12. Also, the authors would like to acknowledge the Bright Spark University of Malaya for the financial support through UM.C/BSU/606/BSP 133(3)-11.

References (68)

  • F. Friedrichs et al.

    Evolutionary tuning of multiple SVM parameters

    Neurocomputing

    (2005)
  • A.C. Lorena et al.

    Evolutionary tuning of SVM parameter values in multiclass problems

    Neurocomputing

    (2008)
  • W.-Z. Lu et al.

    Potential assessment of the “support vector machine” method in forecasting ambient air pollutant trends

    Chemosphere

    (2005)
  • Tirusew Asefa et al.

    Multi-time scale stream flow predictions: the support vector machines approach

    J Hydrology

    (2006)
  • Y. Ji et al.

    Multitask multiclass support vector machines: model and experiments

    Pattern Recognit

    (2013)
  • S. Rajasekaran et al.

    Support vector regression methodology for storm surge predictions

    Ocean Eng

    (2008)
  • Haiqin Yang et al.

    Localized support vector regression for time series prediction

    Neurocomputing

    (2009)
  • K.-P. Wu et al.

    Choosing the kernel parameters for support vector machines by the inter-cluster distance in the feature space

    Pattern Recognit

    (2009)
  • Shahaboddin Shamshirband et al.

    Wind turbine power coefficient estimation by soft computing methodologies: comparative study

    Energy Convers Manag

    (2014)
  • J. Adamowski et al.

    A wavelet neural network conjunction model for groundwater level forecasting

    J Hydrology

    (2011)
  • Z. Peng et al.

    Application of the wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography

    Mech Syst Signal Process

    (2004)
  • A.M. Kalteh

    Monthly river flow forecasting using artificial neural network and support vector regression models coupled with wavelet transform

    Comput Geosciences

    (April 2013)
  • Martin Casdagli

    Nonlinear prediction of chaotic time series

    Phys D

    (1989)
  • M. Marcellino et al.

    A comparison of direct and iterated multistep AR methods for forecasting macroeconomic time series

    J Econ

    (2006)
  • G. Chevillon et al.

    Non-parametric direct multi-step estimation for forecasting economic processes

    Int J Forecast

    (2005)
  • R. Ramanathan et al.

    J Forecast

    (1997)
  • K. Wojdyga

    An influence of weather conditions on heat demand in district heating systems

    Energy Build

    (2008)
  • J.S. Armstrong et al.

    Error measures for generalizing about forecasting methods: empirical comparisons

    Int J Forecast

    (1992)
  • N. Cordeschi et al.

    Energy-saving self-configuring networked data centers

    Comp Netw

    (2013)
  • S. Frederiksen et al.

    District heating and cooling

    (2013)
  • C. Reihav et al.

    Profitability of sparse district heating

    Appl Energy

    (2008)
  • Population Division, Department of Economic and Social affairs, United Nations. World urbanization prospects: the 2009...
  • Population division, department of economic and social affairs

    (2010)
  • Connolly D., Lund H., Mathiesen B.V., Werner S., Möller B., Persson U., et al. Heat roadmap Europe: combining district...
  • Cited by (85)

    View all citing articles on Scopus
    View full text