Page 256 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

248 CHAPTER 12 Computational Intelligence in the Time

relevance, the aspect has been widely addressed in the related literature [4] that,
however, leaves major investigation areas open.
The key point of harvesting is to maximize energy acquisition that, scavenged
from an available source of power, is stored either in an energy accumulator (e.g.,
a battery or a supercapacitor) or directly consumed by the electronic system.
Likewise, the major goal of energy management is to intelligently control energy
consumption by acting at the hardware and software levels, with most sophisticated
decision strategies based both on available and forecasted energy availability.
Machine learning and, more in general, computational intelligence techniques such
as adaptive models for time-series prediction and change-point detection are suitable
methods to be considered here, since physical descriptions about the harvestable
energy is unavailable due to time variance of the environment and accurate informa-
tion about the current power consumption is available through measurements only.

3.1 ENERGY HARVESTING

In CPSs energy can be harvested by relying on different technologies, for example,
from photovoltaic or Peltier cells to wind and flow turbines and, again, by relying on
piezoelectric solutions. The optimal solution for a generic CPS application depends
on the availabledand harvestabledenergy, no matter whether the available power is
high or not. “We get what we have, and that has to be enough,” says an old leitmotiv.
However, of particular relevance for the high density power harvestable are solutions
coming from small photovoltaic cells. They can be both deployed outdoor and
indoor, with polymer flexible cells that, though less performing compared to the
crystalline or amorphous counterparts, are foldable and assume any shape to fit
with the available surface. This flexibility makes these harvesting solutions
appealing with IoT since the cell can be designed to be deployed directly on the
target object. Photovoltaic cellsdbut the same can be stated for wind and flow
turbinesdcan greatly take advantage of intelligent solutions in the sense that we
can maximize the acquired power through adaptation mechanisms. Adaptation is
needed every time the energy source, here the light, provides a power density that
evolves with time, for example, due the presence of clouds, dust, or water drops
on the cell surface, as well as due to varying incidence angles. v p is defined as the
controllable voltage imposed at the photovoltaic cell; the harvestable power
depends, for example, on the particular effective solar radiation as per Fig. 12.1A,
where curve A is associated with a stronger power availability compared with
case B. v p can be controlled over time by means of a Control Power Transfer Module
(refer to Fig. 12.1B) that grants the harvester to maximize the extracted power to be
sent to the storage mean. Details can be found in Ref. [4]. Adaptation can be seen as
an online learning procedure where the optimal controlling parameters at time kþ1
can be achieved through a gradient ascent algorithm as

dði p v p Þ dði p Þ
v p ðk þ 1Þ¼ v p ðkÞþ g ¼ v p ðkÞþ g i p þ v p (12.1)
dv p dv p