Final Report Summary - COMMIT-NRG (Computational intelligence methods in time-series analysis and forecasting with application to energy management systems)
The work of Dr Pedro Ferreira, during the project COMMIT-NRG (European Reintegration Grant (ERG) No. 239451, May 2009 - April 2012) was centred in two main goals:
i) to develop electricity demand predictive models to be used by the Portuguese Power-Grid Company (PPGC) Rede Electrica Nacional (REN), and
ii) to improve the efficiency of existing HVAC systems through the use of model-based predictive control (MBPC).
In order to obtain these goals, research was also conducted in more general aspects of the application of computational intelligence for the modelling and analysis of time-series.
A brief summary of each aspect follows:
Electricity load demand models
This work was also funded by a contract with PPGC. The Portuguese power grid company wanted to improve the accuracy of the electricity load demand (ELD) forecast within an horizon of 24 to 48 hours, in order to identify the need of reserves to be allocated in the Iberian market. To accomplish this, the evolution of ELD over a prediction horizon of at least 48 hours was required.
To address this, artificial neural networks (ANN) were employed as the modelling tool for the electricity load demand time-series. They are used as on-step-ahead predictors that, every time a forecast is required, are executed and simulated in order to build a complete hourly forecast up to 48 hours ahead. The models were designed using information from several years and compute the forecast at a given time on the basis of the load demand within the previous week. In order to improve the accuracy, the models take into account information as the day of the week and holidays. As there are many design possibilities for the models considered, an evolutionary computing approach, namely multi-objective genetic algorithms, were used to search for specific model instances that would be capable of achieving specified design and accuracy goals. Finally, one of such models was selected by means of its performance on a set of data specifically selected for that purpose. In order to provide a comparison baseline, a predictive nearest-neighbour approach was also studied and implemented. This is an intuitive method with good predictive ability that should be exceeded by any method to be used in practice.
An information system has been developed in order to accommodate, execute and maintain information related to the execution of multiple generic multi-step predictive models. Both this system and the select ANN model, plus the nearest neighbour predictive approach, were installed at the PPGC operational facilities and are in execution since may 2010.
A detailed analysis of 22 months of operation concluded that on average the absolute error is 3 % for a prediction horizon of 24 hours, and 4 % for a 48 hours horizon.
Predicting the electricity load demand with such accuracy, enables institutions or companies at regional or country-wide scale, to optimise the management of electricity production plants and the operations of buying electricity as/when needed, therefore achieving important economical savings.
It was also concluded that further improvements may be attained if the atmospheric temperature is taken into account by the predictive models. This is the subject of ongoing research to complement the methodologies implemented in this project.
Model-based predictive control (MBPC) of HVAC systems
This work was also nationally funded by project FCT PTDC/ENR/73345/2006 and Ceratonia 2008 award. Using the same combination of modelling techniques as described above, the goal in this project component was to propose and formalise a predictive control strategy for the HVAC systems used in buildings. The objectives of such strategy are the maintenance of the thermal comfort of the building occupants and the minimisation of the energy consumption to achieve that.
For that purpose, three areas of one building of the University of Algarve were equipped with wireless sensor networks in order to monitor the several aspects of various rooms: air temperature and humidity, mean radiant temperature, the state of doors and windows, and the level of activity within the rooms. Additionally, a software layer was developed in order to allow interfacing with the existing HVAC components installed in the building, and a weather station was installed on top of one building. By using the three systems, the researchers and students involved were capable of monitoring the rooms, the external weather, and of controlling the existing HVAC units.
By using the model identification techniques already mentioned above and the sensor information gathered by the WSNs and weather station, a number of predictive models were identified, which enable the possibility of obtaining accurate real-time localised weather forecasts within a 4 hr horizon. Additionally, predictive models were developed for the rooms which enabled the possibility of simulating the rooms thermal response to the combination of external weather and selected control HVAC actions. A final model was used to assess the occupants thermal comfort: the widely accepted predicted mean vote (PMV) index. This index relates the room environment measurements to the percentage of occupants that feel comfortable within the room.
By using all the models in the correct sequence, a model-based predictive control algorithm was formalised and implemented that, in real-time, searches for the best future trajectory of HVAC control actions that maintain the percentage of dissatisfied occupants below 5 % and minimises the energy spent to achieve that. Once the best trajectory is found, the relevant control action is applied to the system, and the algorithm waits for the next execution instant, where the methods will be repeated.
Several experiments were conducted, both in winter and summer conditions. When compared to the normal HVAC operation, energy savings ranging from 41 % to 77 % were observed. We can therefore extrapolate that, for the type of HVAC systems that were used, and for a hot climate similar to the region of Algarve, important savings in energy, in the order of 50 %, are to be expected using this methodology.
By considering that in the European Union (EU) primary energy consumption in buildings represents about 40 % of the total energy consumption, and, with variations from country to country, half of this energy is spent for indoor climate conditioning, significant savings can be achieved by using HVAC control and managements systems like the one implemented.
Ongoing research is being conducted in order to transform the methods that resulted from the this research into a product that may be easily integrated with existing HVAC systems in buildings.
i) to develop electricity demand predictive models to be used by the Portuguese Power-Grid Company (PPGC) Rede Electrica Nacional (REN), and
ii) to improve the efficiency of existing HVAC systems through the use of model-based predictive control (MBPC).
In order to obtain these goals, research was also conducted in more general aspects of the application of computational intelligence for the modelling and analysis of time-series.
A brief summary of each aspect follows:
Electricity load demand models
This work was also funded by a contract with PPGC. The Portuguese power grid company wanted to improve the accuracy of the electricity load demand (ELD) forecast within an horizon of 24 to 48 hours, in order to identify the need of reserves to be allocated in the Iberian market. To accomplish this, the evolution of ELD over a prediction horizon of at least 48 hours was required.
To address this, artificial neural networks (ANN) were employed as the modelling tool for the electricity load demand time-series. They are used as on-step-ahead predictors that, every time a forecast is required, are executed and simulated in order to build a complete hourly forecast up to 48 hours ahead. The models were designed using information from several years and compute the forecast at a given time on the basis of the load demand within the previous week. In order to improve the accuracy, the models take into account information as the day of the week and holidays. As there are many design possibilities for the models considered, an evolutionary computing approach, namely multi-objective genetic algorithms, were used to search for specific model instances that would be capable of achieving specified design and accuracy goals. Finally, one of such models was selected by means of its performance on a set of data specifically selected for that purpose. In order to provide a comparison baseline, a predictive nearest-neighbour approach was also studied and implemented. This is an intuitive method with good predictive ability that should be exceeded by any method to be used in practice.
An information system has been developed in order to accommodate, execute and maintain information related to the execution of multiple generic multi-step predictive models. Both this system and the select ANN model, plus the nearest neighbour predictive approach, were installed at the PPGC operational facilities and are in execution since may 2010.
A detailed analysis of 22 months of operation concluded that on average the absolute error is 3 % for a prediction horizon of 24 hours, and 4 % for a 48 hours horizon.
Predicting the electricity load demand with such accuracy, enables institutions or companies at regional or country-wide scale, to optimise the management of electricity production plants and the operations of buying electricity as/when needed, therefore achieving important economical savings.
It was also concluded that further improvements may be attained if the atmospheric temperature is taken into account by the predictive models. This is the subject of ongoing research to complement the methodologies implemented in this project.
Model-based predictive control (MBPC) of HVAC systems
This work was also nationally funded by project FCT PTDC/ENR/73345/2006 and Ceratonia 2008 award. Using the same combination of modelling techniques as described above, the goal in this project component was to propose and formalise a predictive control strategy for the HVAC systems used in buildings. The objectives of such strategy are the maintenance of the thermal comfort of the building occupants and the minimisation of the energy consumption to achieve that.
For that purpose, three areas of one building of the University of Algarve were equipped with wireless sensor networks in order to monitor the several aspects of various rooms: air temperature and humidity, mean radiant temperature, the state of doors and windows, and the level of activity within the rooms. Additionally, a software layer was developed in order to allow interfacing with the existing HVAC components installed in the building, and a weather station was installed on top of one building. By using the three systems, the researchers and students involved were capable of monitoring the rooms, the external weather, and of controlling the existing HVAC units.
By using the model identification techniques already mentioned above and the sensor information gathered by the WSNs and weather station, a number of predictive models were identified, which enable the possibility of obtaining accurate real-time localised weather forecasts within a 4 hr horizon. Additionally, predictive models were developed for the rooms which enabled the possibility of simulating the rooms thermal response to the combination of external weather and selected control HVAC actions. A final model was used to assess the occupants thermal comfort: the widely accepted predicted mean vote (PMV) index. This index relates the room environment measurements to the percentage of occupants that feel comfortable within the room.
By using all the models in the correct sequence, a model-based predictive control algorithm was formalised and implemented that, in real-time, searches for the best future trajectory of HVAC control actions that maintain the percentage of dissatisfied occupants below 5 % and minimises the energy spent to achieve that. Once the best trajectory is found, the relevant control action is applied to the system, and the algorithm waits for the next execution instant, where the methods will be repeated.
Several experiments were conducted, both in winter and summer conditions. When compared to the normal HVAC operation, energy savings ranging from 41 % to 77 % were observed. We can therefore extrapolate that, for the type of HVAC systems that were used, and for a hot climate similar to the region of Algarve, important savings in energy, in the order of 50 %, are to be expected using this methodology.
By considering that in the European Union (EU) primary energy consumption in buildings represents about 40 % of the total energy consumption, and, with variations from country to country, half of this energy is spent for indoor climate conditioning, significant savings can be achieved by using HVAC control and managements systems like the one implemented.
Ongoing research is being conducted in order to transform the methods that resulted from the this research into a product that may be easily integrated with existing HVAC systems in buildings.