predicting solar generation from weather forecasts using machine learning

Post on 19-Jan-2016

48 Views

Category:

Documents

2 Downloads

Preview:

Click to see full reader

DESCRIPTION

Predicting Solar Generation from Weather Forecasts Using Machine Learning. Navin Sharma, Pranshu Sharma, David Irwin , and Prashant Shenoy. Harvesting Examples. Perpetual Sensor Networks Run forever off harvested energy[ EWSN 2009 ] Off-the-grid infrastructure Power cellular towers & ATM - PowerPoint PPT Presentation

TRANSCRIPT

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Predicting Solar Generation from Weather Forecasts Using

Machine Learning

Navin Sharma, Pranshu Sharma,David Irwin, and Prashant Shenoy

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Harvesting Examples

Perpetual Sensor Networks Run forever off harvested energy [EWSN 2009]

Off-the-grid infrastructure Power cellular towers & ATM

Smart homes and smart cities Use on-site solar & wind power [BuildSys

2011]

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Renewables are Intermittent

Example: Solar shows significant variation

Nearly no energy

How much energy will we harvest today?

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Predictions are Important

Better predictions == Better performance Examples:

Smart homes [BuildSys 2011] Reduce utility bill by 2.7X Eliminate peak power demands

Sensor Network [SECON 2010] Lexicographical sensor network: increases sensing rate by

60% Sensor testbeds: serve 70% more requests

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Prediction Methods

Existing Prediction Methods Past Predicts Future (PPF) Variants of PPF

EWMA [TECS 2007] WCMA [VITAE 2009]

Past Predicts Future Accurate for short time scales (seconds to minutes) Hard to predict at medium time scales (hours to

days)

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Problem Statement

How can we statistically predictsolar harvesting ?

Approach: Leverage weather forecast to predict solar energy Use statistical power of machine learning

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Outline

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Forecast-based Predictions

Idea for using weather forecasts PPF accurate for constant weather Forecasts also predict significant weather

changes

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Methodology

Analyze Weather Data Forecast data from National Weather Service

Formulate Forecast Solar Intensity Model Use machine learning regression techniques Solar Intensity = F (time, multiple weather

parameters)

Derive Solar Intensity Solar Energy Model Empirically from our solar panel deployment

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Data Analysis

Solar intensity exhibits strong (but not perfect) correlation with sky cover, humidity, and precipitation

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Data Analysis

Solar intensity exhibits no correlation with wind speed, but weak correlation with temperature

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Prediction Technique

ML Regression Techniques Training data set to find regression coefficients Testing data set to verify the model’s accuracy

Our data set Training data set: First 8 months of 2010 Testing data set: Next 2 months of 2010

What to predict? Solar intensity at noon Based on 3-hr weather forecast at 9 AM

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Support Vector Machines Support Vector Machine (SVM)

Used for classification & regression Independent of input space dimensionality Resistant to overfitting

Kernel Function Maps data from low-dimensional input space to high-

dimensional feature space Common Kernels

Linear kernel Polynomial kernel Radial Basis Function (RBF)

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

SVM Regression: Steps

Step 1: Data Preparation Normalize to zero mean and unit variance

Step 2: Kernel Selection RBF performs better than linear & polynomial Grid search to find optimal parameters Optimal parameters:

cost (soft margin parameter) = 256 γ (Gaussian function parameter) = 0.015625 ε (loss function parameter) = 0.001953125

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

SVM with RBF Kernel

Average prediction error: 22 %

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Dimensionality Reduction

Redundant Information Reduces prediction accuracy

Principal Component Analysis (PCA) Correlated variables uncorrelated variables Uncorrelated variables called principal components Choose first 4 PCs with first 4 (highest) Eigen values

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

SVM with RBF Kernel

Reducing dimensions from 7 to 4 reducesprediction error from 22 % to 2 %

4-dimensions

7-dimensions

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Comparison with Cloudy Model

SVM-RBF with 4 dimensions predicts 27 % better than cloudy-forecast

SVM-RBFCloudy-forecast

Cloudy-forecast: Sky cover based empirical model for solar prediction [SECON 2010]

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Intensity Energy Model Solar power from solar intensity

Depends on solar panel characteristics Panel orientation & surrounding environments Empirically derived for a particular setup

Our solar panel deployment Kyocera KC65T Solar Panel Power = 0.0444 * Intensity - 2.65

Accurate to within 2.5 % of actual harvesting

UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTS ASSACHUSETTS AAMHERST • MHERST • Department of Computer Science Department of Computer Science • 2011 • 2011

Conclusions

Weather forecasts can improve prediction accuracy See dramatic weather changes before they occur Facilitates better planning ML statistical models work well

Future Work Design a better kernel function Hybrid Prediction: use a combination of past & forecast Apply to wind and wind gust

top related