Most of the highly complex systems in our modern world consist of a great number of electronic components. All these components need electrical energy to run and, if the number of components is scaled up, also the power consumption increases rapidly.
As this is a very big cost factor, component suppliers have countered this by building individual power saving techniques into their products. These techniques mostly monitor the energy consumption of one individual component of the whole system and decide, depending on the individual usage requirements, on switching to power saving modes or slower operation modes.
Take a cell phone network as an example: It consists of a huge amount of individual components that are of various sizes and power requirements. There are components in the backbone of the network, that can handle the switching of thousands of users and telephone calls at the same time. Then, in the direct connection to the users of the network, there are cell towers, which in again consist of a logical switching component and several power amplifiers and transmitters for actual transmission of the data. Those components of course can be regulated individually, for example the power amplifier can increase and decrease the transmit power, that is fed to the antenna. Since this is the component in a mobile base station, that uses the most power, a lot of energy can be saved by intelligent control of this amplifier alone.
If there are now various neighboring cell towers in an area, that can each individually control their power, interference effects may increase, where the power saving strategy of one cell tower undermines all energy saving effects of another tower. In such a complex scenario it is therefore necessary to analyze the whole network as a highly complex system with thousands of individual components, that each have their individual capability of power management.
The idea is now to incorporate all the knowledge, collected from each component in the system, analyze this data and derive decisions, which can help decrease the overall power consumption.
To achieve this, numerous components are required: First of all, an intelligent and versatile data collection has to be implemented in order to enable the system to make a informed and global decision. Second, efficient learning algorithms have to be employed to analyze this huge amount of data, since manual tuning is no longer possible for a human operator. Third, effective decision making strategies and mechanisms have to be implemented in the system, that on the one hand use already implemented monitoring and data collection facilities and on the other hand can remotely or locally control the power saving techniques already built in to the systems components.
One central element of the Smart Energy Management System is the advanced data fusion. At the same time, the data fusion component must provide both the actual state of the whole system and all available energy saving features of the attached components. The state of the system is on the one hand monitored and reported by each individual component but also global measurements of the whole network (for example the network load) are important.
The actual power consumption or load management and monitoring of specific components is already implemented in many system components like power supplies, control units or transceivers (considering a distributed network infrastructure). Using existing measurement tools like smart meters or commercial power monitoring devices can extend or upgrade these capabilities. The actual system load and state must also be observed and reported. Therefore, in network components the SNMP or Netflow protocol exists. All these information packets must securely be transferred to the main data fusion core.
The fusion core has to be able to handle various different protocols and information sources. To achieve this an extensible and powerful architecture is needed to receive and store all incoming data packets. Then, the core has to process and integrate the data into a consistent model of the system. The data fusion core must be able to continuously handle the incoming data and update the system state model which is used for the autonomous energy saving profile calculation.
Besides the power consumption and system load the data fusion core provides information of energy saving capabilities and their actual states of all connected components. It is important that the core "knows" beside simple energy saving features also advanced technologies which for example might affect both network bandwidth and the energy consumption of the power amplifier. Only using higher level techniques which combine advanced control options and optimization techniques of the distributed system will have enough potential for a significant saving of energy.
The second central element of the whole system design is the optimization learning. This means that data, prepared and partly preprocessed by the data collection core, will be analyzed as a whole. This can be done by employing a model and fitting it to the available data. Then from this modeled system, the optimal parameters can be deduced. Unfortunately, in very complex systems, it is often unfeasible to even approximately model the whole system. Therefore also model-free learning and data analysis techniques should be employed.
Another hurdle to jump is the amount of data. If the system will be deployed into a live mobile network for example, an incredible amount of data will be produced and has to be processed. Recent research in big data processing shows promising results on analysis and processing strategies of very big amounts of data. Often, learning algorithms are not yet tailored to run in a data processing optimized framework. Therefore, current and future learning algorithms and analysis techniques have to be adapted to work in this environment of real time data collection and power save decision making.
A live system is very vulnerable. Every disturbance or connection loss would mean extra cost for the mobile provider in communications networks for example. This is why the learning techniques have to be designed to prevent such a unfavourable decisions. To keep the management of the system in the control of the provider, a learning system could also be designed in such a way, that it just analyzes the data and is only allowed to make non-intrusive decisions. All possible harmful decisions are delayed and to be checked by the operator first. Another strategy would be to not control the system at all, but to just propose control strategies that would reduce the energy consumption. The actual execution of those control strategies still remains with the operator.