DREAM: Dynamic real-time control of energy streams in buildings
|Funding||ICTRegie, NWO en STW|
|Duration||2011-04-01 ~ 2014-04-01|
|Staff||Johann Hurink, Albert Molderink, Gerard Smit, and Hermen Toersche|
These goals are in line with the EU targets for 2020: 20% reduction in emissions compared to 1990 levels; 20% share of renewable energies in overall EU energy consumption; and 20% savings in EU energy consumption compared to projections. One of the problems of integrating renewable energy resources in the generation process, however, is that they increase the dynamism of the electricity system. Many of the generators using renewable sources only provide energy at certain times of the day (e.g. a PV produces more when there is direct sunlight and windmills generate only when there is wind) and they often cannot be controlled. A substantial CO2 reduction can be achieved by using the renewable sources in a more efficient way e.g. by temporary storing energy or by clever peak shaving. It is generally agreed that ICT technology plays an important role to reach such CO2 reductions. In this proposal we focus on ICT technologies that can be employed in buildings that are interconnected in a small electricity micro-grid of approx. 150 buildings, which is the typical size of the low-voltage electricity grid behind a local transformer.
Step 1: prediction of energy patterns
In the first step of the three-step optimization method a prediction of the energy demand and the electricity production potential is made for each building individually. This is done for every building separately, because every building has a different power profile due to different local generators, different appliances used in the building and different behaviour of people working or living in these buildings. In the SFEER project we concluded that it is possible to observe regularities and to predict to some extend the energy demand of individual buildings.This local information forms the base to determine the local electricity production potential and possible scheduling freedom of the micro-generators, buffers and appliances.
Next to the energy profiles of individual buildings in the micro-grid, the energy profile of generators not directly connected to buildings like local wind-mills and local biogas installations within the micro-grid are predicted in step 1.
Step 2 global scheduling of energy streams
In the second step the local production and consumption potential is used to compute an integral planning of the micro-grid. This planning specifies the preferred power profile of all individual buildings. It is calculated 24 hours ahead, based on local (building specific) and global objectives (e.g. APX electricity wholesale prices), and is based on the information the global controller receives from step 1 such as e.g. micro-generator characteristics and (remaining as well as maximum) heat buffer and battery capacity. The global planning does not have to be restricted to the micro-grid. For this project we assume that the profiles computed in step 2 are considered as preferred power profile for the individual building. A power profile specifies for each time interval during the next 24 hours (e.g. 15 minute periods) the balance between electricity consumption and production. This is specified by minimum, maximum and preferred energy consumption per time interval
The main idea of the proposed planning method is to organize the buildings of the grid in a hierarchical way; i.e. as a tree where the leaves represent the buildings. Each node in this hierarchical structure represents a subset of the grid (e.g. a building, a neighbourhood or a complete city) and the planner within a node only communicates with the planners in the nodes above and below him. Using this hierarchical structure, an iterative method has been developed, that gives an approximation of the global objective and that scales with the grid size. Within this method, the planners can steer their sub-grid by sending (artificial) cost prices of electricity to the planners below. These prices are adapted based on signals the planner gets from the planner from the level above. On the leaf level, the local (house) controllers schedule their building such that energy consumption is shifted to periods with low prices and energy production to periods with high prices. By iteratively sending different prices (based on deviation from the target value of the fleet), the controllers in the houses reschedule their buildings, resulting in an aggregated planning of the whole grid which matches better the global objective. As the global planning problem has been shown to be NP-complete in the strong sense we may not expect to achieve efficient methods solving the global scheduling problem to optimality.
Step 3: local control
In step 3 the control system has to control the supply and demand within a building such that the total energy consumption of the building is close to the preferred power profile. This last step is performed by a local real-time controller which decides when to switch on/off appliances, when to charge batteries, or when to store heat in the heat buffer, etc. Whereas the first two steps can be done off-line, the devices need to be (on-line) controlled in real-time.
On-line decisions have to be made to switch devices on or off. These decisions should not only be based on the preferred planning but should also take into account the current situation in the building (e.g. new devices may (request to) start, the temperature in the fridge may change after filling it with new goods). In this way, the controller has to take into account local constraints resulting from the concrete situation in the current time period, but should also estimate the influence of the made decisions for future time periods.
We propose to base the real-time control on cost functions for every device (generators, consumers, storage) and to take as goal to minimize total costs. Using these cost functions not only the preferred plan but also the priorities of the residents and the possible incentive based on which they allow some discomfort, can be taken into account. In this way, the local controller has to find a balance between these possibly conflicting objectives, resulting in schedules deviating from the preferred planning. As the local controller has no knowledge of the state of the overall system, but can only take into account the steering signals (e.g. the preferred power profile) it has received from a global controller, communication with the global controller about these estimated deviations from the preferred power profile in future time periods may be used to better judge the influence these deviations have for the performance of the overall system
The proposed controller should also work if there is no (two-way) communication with a global controller (e.g. because there is temporarily no network connection). In this case, the local system can fall back to another (locally) focussed optimization scheme like peak shaving. Although this might not reach a global optimum, still energy efficiency can be improved since flat demand patterns can be supplied more efficiently.
Essential in the sketched approach is that we propose to make schedules for the complete planning horizon of one day (24 hours) and not only for the current time period. This is quite different from smart agent approaches that react in real-time on signal from other agents. In our approach extreme behaviour of the system, for example all agents reacting at the same time in the same way on pricing signals, can be avoided. We expect that this approach works smoother, even in highly dynamic and extreme situations, and can give more guaranties for the performance of the overall system. Therefore, our approach is much more transparent than agent based approaches.
As already mentioned, the predictions 24 hours in advance surely will deviate from the reality and therefore the preferred planning may not be realized. The local controller has to work around these prediction errors and still follow the planning as close as possible. Therefore, the local controller should be able to “look a few hours in the future” and make a short-term planning that is feasible and follows the planning as good as possible. For this short-term planning a short-term prediction can be used. Since more knowledge about the status of the devices is available and the prediction is only for the short-term, these prediction will be much more accurate than the 24 hours ahead predictions.
In step 2 the finally achieved global production planning is based on the local plannings of all buildings. These local plannings were calculated based on the 24 hour demand predictions and on the steering signals in the last iteration of the global planning method. In general, these local plans do not specify the schedules of all devices, but mainly specify the schedules for the individual micro-generators or storage devices and maybe some ‘large’ devices like e.g. washing machines and electrical cars. The other devices are summarized in a sort of base load but are not precisely scheduled. If during the short term planning the reality differs too much from that schedule, the individual micro-generators or storage devices may need to be rescheduled compared to the master schedule. In this way we can cope for prediction errors and the stochastic nature of (e.g. human caused) demand. Furthermore, now also devices with long running times (e.g. washing machine or dish washer) can be scheduled. For this short term planning not only the preferred planning from the global controller but also the comfort level in the building (e.g. temperature in a room or fridge) has to be taken into account.
For realizing a short-term planning, concepts like Model Predictive Control (MPC) or a rolling horizon approach may be developed. Besides the concrete methods and algorithms for calculating the concrete schedules in each iteration of such a method also questions like the re-planning frequency or the length of the time horizon taken into account in the local plannings have to be answered. Furthermore, when the MPC or rolling horizon approach conclude that the short-term planning differs too much from the planning which formed the base of the global planning, the local controller can request a re-planning on a higher level. To control somehow the extra effort resulting from such re-planning steps, good communication and decision protocols between the planners on the different levels of the hierarchical structure have to be developed. Furthermore, since a re-planning should not shuffle the existing plans too much and has to be executed faster than the global planning a day ahead, new methods and algorithms for the re-planning have to be developed. These methods should take into account the hierarchical structure and the should decide on which of the different levels the re-planning should be done. Some deviations may be corrected by only re-planning houses in the neighbourhood, whereas in other cases a re-planning over the whole grid is needed.