Lorenzo talk on GPU computing for cloudiness tracking and prediction

Lorenzo talk on GPU computing for cloudiness tracking and prediction

Tony Lorenzo - IES Renewable Power Forecasting Group

Solar power forecasting, data assimilation, and El Gato.

 

Forecasting partners with regional power supply companies.

  1. Solar variability

Early AM full power as panels track sun

PM variability as clouds move across the sky.

 

  1. Irradiance to power consumption
  2. Clear-Sky Index
    1. Clear-sky expectation

Observation

TEP asked for solar forecasts becasue they saw variability as in issue.

Deployed irradiance sensor network

Network forecasts

Using GEOS 

GHI - conversion of cloudiness to transmittance

Comparing irradiance forecasts errors

their errors are on the order of 150 (clearsky) to 120 (WA-WRF) to 50 (sensor network) W/m2

Satellite imae after optimal interpolation.

Satellite derived GHI estimate

Two conversio nmodels

a semi-empirical model that applies a regression to the data of visible images

a physical model that estimates cloud properties and performs radiative transefer

Nominal 1km resolution Adjusted visible albedo

Using 75km x 82 km area over tucson

Optimal interpoliation

Bayesian technique for minimizing the se distance between the field and observed

best linear unbiased estimated 

Kalman filter 

x_b = x_t + g

g~ N(0,P)

Obs

y=Hx_t +e

e ~ N(0,R)

x_a=x_b + W(y -H_x_b)

W= PH^T (R+HPH^T)^-1

Error covariances P and R

decompose P into a diagonal variance matrix and correlation matrix: P=D^0.5 C D^0.5

prescribe a correlation between the image pixels based on he difference incloudiness to construct C a

Compute D form cloud free training images

Assume observation arros from the sensors are uncorrelated

Used Python on elgato

import numpy as np

import scipy import linalg

import skcuda.linalg as cu

from pycuda import gpuarray

Dask

Numba

Singularity

PyCUDA

scikit-cuda

Sumatra - provenance tracking.

Travis Hardy - data prediction

Ensemble Kalman Filter