====== Forecasting Meeting Minutes for Week of October 18, 2016 ====== ** Present Monday: Austin, Gordon, Brieanna, Jaimie ** ** Present Tuesday: Gordon, Brieanna, Jaimie ** ===== Updates ===== * Ask Sharif to meet with us ===== Meeting with Dr. Kuh ===== ===== Expectations ===== * 3D plot: 2011, 2012 normalized and unnormalized for comparison * LA more variation, but smoother * Moving Average and Decimation/Downsample * D and X from 2011: * X = 8:00-4:00 (input data) * D = 9:00-5:00 (normalized and downsampled, column) * Y = Least Squares Equation = W.T(X) row = X.T(W) column * Unnormalized = D bar, Y bar * Statistical: * Y - sig*Y + m * D - sig*D + m * Zenith: * Y*cos(theta) * D*cos(theta) * W = (XX.T)^(-1)*XD <= XX.T*W = XD * E^2 Averaged = (D-Y)^2 for every time * MSE = 1/m (||D-Y||)^2 * Don't need r bc cancel * Ask Seyed what he did for decimation/downsample * 1) Average and down sample * 2) Take moving average * Overlapping data points: * 8:01-8:10 moving average * Then downsample * Filter then normalize * Normalize then filter * **Normalization: mean then deviation** * Zenith Angle Filter: Zenith angle is from the 5th minute, 5 irradiance is averaged * Bias term * Number of taps is # present + number before * 1 tap = present * 2 tap = present + future (5 mins in our case) * In weight matrix * Row of 1's to the matrix to get bias term * Statistical normalization: -mean/deviation * Standard normalization 0 mean * Estimate will need to be given back after the normalization * 3D plot for both zenith and stat normalization * Plot MSE * y = MSE, x = # taps * Train error goes down, test error decreases until flattens or increases * Check 2012 MSE test against 2011 MSE train * Keep 2011 W * Get new X, D to compute new Y * Flip: 2011 MSE test against 2012 MSE train * Correlation is 3D plot * R=(1/m)XX.T= 10x10 matrix that is symmetric