Math 390 - Machine Learning Techniques Presentation
1. Machine Learning Techniques for Solar Radiation Modeling
Hamza Amhaz Darragh Punch Chase Rosenthal Connor Winton
2. Artificial Neural Networks (ANNs)
Why ANNs?
• ANNs are useful in predicting the most efficient spatial area of a PV system
(Mellit, et al. 2003.)
• Neural networks can be trained to predict results from examples and “learn” from their
mistakes and successes. (Mellit, et al. 2003.)
• How does a child learn not to touch a boiling pot?
3. Artificial Neural Networks (ANNs)
ANNs and PV Systems
• Mellit, et al. define the PV system capacity as follows:
• 𝐶𝐴 =
η 𝐺 𝐴 𝐺 𝐻
𝐿
• 𝐶𝐴: PV system capacity
• η 𝐺: PV system efficiency
• 𝐴 𝐺: PV system physical area
• 𝐻: Avg. daily solar radiation
• 𝐿: Avg. daily energy consumption
Back-propagation Algorithms
4. Artificial Neural Networks (ANNs)
A: PV system spatial area
C: PV system useful capacity
p: predicted
m: measured
(Mellit et al. 2003.)
5. Support Vector Machines (SVMs) and Radial Basis Functions (RBFs)
Why SVMs?
• SVMs are 27% more accurate when predicting solar radiation using weather forecast
metrics vs. existing forecast models (Sharma, et al. 2011.)
• SVMs are superior to past-predicts-future models (PPF) which cannot predict
weather changes
• SVMs require the right kernel function (“probability-changing” piece of a prob.
density function) and parameters, which brings in…
RBFs
• Sharma et al. used SVMs with RBF kernels because of the RBFs’ sparsity property and
their ability to handle non-linearity in data
7. Support Vector Regression (SVR)
Lauret et al. offer an SVM and RBF approach with more emphasis on building an
equation:
𝑘∗ 𝑡 + ℎ =
𝑖=1
𝑛
𝑎𝑖 𝑘 𝑟𝑏𝑓 𝑥𝑖, 𝑥∗ + 𝑏
where 𝑘 𝑟𝑏𝑓 refers to the radial basis covariance function, a special kind of RBF.
Why SVR?
• Results from multi-class SVMs were compared to the results from SVR; SVR results
were overwhelmingly a better representation (Lauret et al. 2015.)
8. Non-Traditional Machine Learning Techniques to Consider
Gaussian Processes (GP)
• Lauret et al. suggest GP as the best annual predictor for Global Horizontal solar
Irradiance (GHI) for the Reunion-Saint Pierre power grid
• Lauret et al. present the following equation to forecast the clear sky index using GP:
𝑘∗ 𝑡 + ℎ =
𝑖=1
𝑛
𝑎𝑖 𝑘 𝑓 𝑥𝑖, 𝑥∗
where 𝑘 𝑓 refers to the squared exponential covariance function.
9. Non-Traditional Machine Learning Techniques to Consider
Numerical Weather Prediction (NWP)
• NWP is not explicitly a machine learning technique.
• Since our atmosphere is a fluid, NWP samples the current state of the fluid.
• NWP then uses fluid dynamics to predict the state of the fluid in the future.
Why NWP?
• Lauret et al. cite papers that suggest that NWP models are the most accurate when
trying to forecast weather conditions 6+ hours into the future.
• A forecast of greater scope such as this is particularly relevant to Intel’s investors
who are seeking information to support the idea of building a long-term solar
farm in Arizona.
10. Works Cited
[1] A. Mellit et al., “Modelling of Sizing the Photovoltaic System Parameters Using
Artificial Neural Network,” 2003 IEEE Control Applications Conf., 2003, pp. 353-357.
[2] N. Sharma et al., “Predicting Solar Generation from Weather Forecasts Using
Machine Learning,” 2011 Smart Grid Communications Int. Conf., Brussels, 2011,
pp. 528-533.
[3] P. Lauret et al. A benchmarking of machine learning techniques for solar radiation
forecasting in an insular context. Solar Energy, Elsevier, 2015, pp.00.