This talk covers the application of machine learning techniques for energy applications, in particular for modeling solar radiation. The first part explores meta-heuristic search algorithms and envisioned their application for designing distributed, self-organizing control systems using evolutionary algorithms. The second part gives an introduction to solar radiation modeling and shows how neural networks can be used to artificial neural networks to learn the correlation of input parameters such as latitude, longitude, temperature, humidity, month, day, hour to predict global and diffuse solar radiation.
Machine Learning Techniques for the Smart Grid – Modeling of Solar Energy using AI
1. Machine Learning Techniques for the
Smart Grid – Modeling of Solar Energy
using AI
Networked and Embedded Systems
Professor Dr. Wilfried Elmenreich
Dr. Tamer Khatib
2. Overview
• Scope of this tutorial
• Meta-heuristic search algorithms
• Artificial neural networks
• Modeling of solar radiation
Modeling extraterrestrial and terrestrial solar radiation
Clear sky model
Satellite based models
Sky transmittance-based models
Ground meteorological measurement based model
ANN-based modeling of solar radiation
3. • Automated planning and scheduling
• Machine learning
• Natural language processing
• Perception
• Robotics
• Social intelligence
• Creativity
• Artificial general intelligence
Artificial Intelligence Areas
Image soruce: Creative Commons, Wikipedia
4. • Automated planning and scheduling
• Machine learning
• Natural language processing
• Perception
• Robotics
• Social intelligence
• Creativity
• Artificial general intelligence
Artificial Intelligence Techniques
Image soruce: Creative Commons, Wikipedia
6. • For optimization problems
• Etymology:
– Meta – upper level
– Heuristic – to find
– Heuristic = deterministic
– Meta-heuristic = utilizing randomization in search
• So it is “only” for search problems ?
Every engineering or design challenges can be formulated into a search
problem over a solution space
• Solution space can be particular large and multi-dimensional
– Standard optimization algorithms don’t finish in acceptable time
– Need for meta-heuristic
Meta-heuristic search algorithms
7. Overview on Search Techniques
• Metaheuristics = Guided random search techniques
8. • Metaheuristics are strategies that guide the search process
• Goal is to efficiently explore the search space to find
(near-)optimal solutions
• No single technique
• Metaheuristic algorithms are approximate and typically
non-deterministic
• Metaheuristic algorithms might fail by getting trapped in
confined and deceptive areas of the search space
• Metaheuristics are typically not problem-specific
Properties of Meta-heuristic Search
Algorithms
9. • Trajectory methods
– Basic Idea: Iterative improvement
– Simulated annealing (Scott Kirkpatrick, C. Daniel Gelatt and Mario P. Vecchi, 1983)
– Tabu search (Fred Glover, 1986)
– Variable neighborhood search (Mladenovic, Hansen, 1997)
Meta-heuristic Search Algorithms (1)
x1
x2 x3
X4
X5
12. Searching for Rules
• Simulation of target system as
playground
• Evolvable model of local behavior
(e.g., fuzzy rules, ANN)
• Define goal via fitness function (e.g.,
maximize throughput in a network)
• Run evolutionary algorithm to derive
local rules that fulfill the given goal
System model
Goals (fitness function)
Simulation
Explore
solutions
Evaluate &
Iterate
Analyze results
13. System architecture
Building Self-Organizing Systems 13
Wilfried Elmenreich
6 major components:
task description, simulation setup, interaction
interface, evolvable decision unit, objective function,
search algorithm
14. Agent behavior to be evolved
• Controls the agents of the SOS
• Processes inputs (from sensors) and produces output (to
actuators)
• Must be evolvable
– Mutation
– Recombination
• We cannot easily do this with an algorithm represented in C
code…
Agent
Control System
„Agent‘s Brain“
15. Artificial Neural Networks
• Each neuron sums up the
weighted outputs of the
other connected neurons
• The output of the neuron
is the result of an activation function
(e.g. step, sigmoid function) applied to this sum
• Neural networks are distinguished by their connection structure
– Feed forward connections (layered)
– Recursive (Ouput neurons feed back to input layer)
– Fully meshed
17. Framework for Evolutionary Design
• FREVO (Framework for Evolutionary Design)
• Modular Java tool allowing fast simulation and evolution
• FREVO defines flexible components for
– Controller representation
– Problem specification
– Optimizer
18. Giving FREVO a Problem
• Basically, we need a simulation of the problem
• Interface for input/output connections to the agents
– E.g. for the public goods game:
– Your input last round
– Your revenue
• Feedback from a simulation run -> fitness value
• FREVO source code and simple tutorial for a new problem at
http://frevo.sourceforge.net
20. Application example
• Modeling of solar radiation
Modeling extraterrestrial and terrestrial solar radiation
Clear sky model
Satellite based models
Sky transmittance-based models
Ground meteorological measurement based model
ANN Based modeling of solar radiation
21. • Solar energy is part of the sun’s energy which falls at the earth’s surface. It
can be harnessed, to heat water or to move electrons in a solar cell.
• Solar radiation data provide information on sun’s potential in a specific
location. These data are very important for designing solar energy systems.
• Due to the high cost and installation difficulties in measuring devices, these
data aren't always available. thus, alternative prediction ways are needed.
Preface: Solar energy
22. How big is solar energy ?
Source: Boyle, G. 2004. Renewable Energy. OXFORD..
23. Modeling of extraterrestrial solar radiation
• The Sun emits radiant energy in an amount that is a function of its
temperature. Blackbody model can be used to describe how much
radiation the sun emits. A blackbody is defined to be a perfect emitter as
well as a perfect absorber
• The wavelengths emitted by a blackbody depend on its temperature as
described by Planck’s law:
𝐸𝜆 =
3.74×1010
λ5[𝑒
14.4
𝜆𝑇
−1
]
Where,
Eλ is the emissive power per area (W/m2 μm),
T is the absolute temperature of the body (K),
λ is the wavelength (μm).
24. Modeling of extraterrestrial solar radiation
• To calculate the daily extraterrestrial solar radiation on the top of the
atmosphere, the path that the earth rotates around the sun must be
considered.
• The eccentricity of the ellipse is small and the orbit is, in fact, quite nearly
circular. Therefore, the extraterrestrial solar radiation in W/m2 can be
described as,
𝐼 𝑜 = 1367 ×
𝑅 𝑎𝑣
𝑅
2
where
Rav is the mean sun-earth distance
R is the actual sun-earth distance depending on the day of the year
• After all, the daily extraterrestrial solar radiation can be given as follows,
𝐼 𝑜 = 1367[1 + 0.034 cos
360𝑛
365
]
25. Modeling of terrestrial solar radiation
• Attenuation of incoming radiation is a function of the distance
that the beam has to travel through the atmosphere, which is
easily calculable, as well as factors such as dust, air pollution,
atmospheric water vapor, clouds, and turbidity
26. Modeling of terrestrial solar radiation
• There are many theories for modeling terrestrial solar radiation,
Clear sky model
Satellite based model
Environmental measurement based model
Ground meteorological measurement based model
27. Clear sky model
• Beam radiation at the surface can exceed 70% of the extraterrestrial flux
• Constant and uniform attenuation factor is assumed
• Isotropic model is assumed
36. Number of neurons in the hidden layer
• If a low number of hidden neurons are used, under fitting may occur and
this will cause high training and generalization error while over fitting and
high variance may occur when the hidden layer consist of a large number
of hidden neurons.
• Usually the number of hidden nodes can be obtained by using some rules of
thumb. For example,
• the hidden layer’s neurons have to be somewhere between the input layer
size and the output layer size.
• the hidden layer will never require more than twice the number of the
inputs.
• the number of hidden nodes are 2/3 or (70%-90%) of the number of input
nodes.
• In addition, it has been recommended that by adding the number of the
input to the number of the output and multiply the result by (2/3), the
number of the hidden nodes can be achieved.
38. Summary
• Artificial Intelligence algorithms are complex algorithms to
handle complex problems
• Simple, deconstructable problems (given network, linear
composable power flows) -> standard algorithms
• Complex problems (many variables, open questions such as
network structure) -> complex algorithms
• We covered:
– Evolutionary algorithms
– Artificial neural networks
– Neural network application for modeling of solar radiation
40. Further Links
• Video: 6 minute introduction to FREVO: http://youtu.be/1wTyozYGG4I
• Download FREVO (open source): http://frevo.sourceforge.net
• A. Sobe, I. Fehérvári, and W. Elmenreich. FREVO: A tool for evolving and
evaluating self-organizing systems. In Proceedings of the 1st International
Workshop on Evaluation for Self-Adaptive and Self-Organizing Systems,
Lyon, France, September 2012.
• I. Fehervari and W. Elmenreich. Evolution as a tool to design self-
organizing systems. In Self-Organizing Systems, volume LNCS 8221, pages
139–144. Springer Verlag, 2014.
• T. Khatib, A Mohamed, K Sopian. A review of solar energy modeling
techniques. J. of Renewable & Sustainable Energy Reviews. 2012.16(5):
2864-2869.
• T. Khatib, A. Mohamed, K. Sopian, M. Mahmoud. Assessment of Artificial
Neural Networks for Hourly Solar Radiation Prediction. J. of Photoenergy.
2012. 2012(ID 946890):1-7.