Wind (BLC6) prediction from the previous (5BLC) five days data wind by Kohonen Neural Networks.

After the row of variable labels, the first and second rows are for the KNN training along 10000 epochs, in the learning period 1975–94 (7300 days), supervised by the winds of the 6th days. The third and fourth rows are for the 6th day wind (real) prediction in 1995 by the above training. The fifth and sixth rows are for the 6th day wind (real) prediction in 1995 by similar training but supervised by the winds of the 5th days. The supervising day is indicated by sup, sx is the times the learned matrix has been smoothed with sm = 2/3, coin is the number of total (wind direction plus vorticity) coincidences between observed and predicted winds. Next columns to the right are, the first set (δV0 nw-w δα) for the number of days, δV0, having the same vorticity for the observed and the predicted winds, the second set when the vorticity difference is 1, δV1, and the third set when the difference is 2, δV2, (A to C or C to A). The other two columns of each set are the number of days changing from no-wind to wind (nw-w), or vice versa, and the average angle between the observed and predicted winds (δα). The last three columns are the average localization error for the 5BLC vectors <δ_loc>, the prediction error for the BLC6 vectors <δ_pred>, and the same error but approaching the matrix components to 1, 0 or −1, to get the closest possible observed vector <δ_pred >1.