Conclusions

          The exploratory graphics and the gradient analysis provided us with a little bit of insight to the relationships between the WRF model's errors and the predicted weather attributes. Also by applying the RandomForest algorithm to predict the model's bias we gain a noticeable improvement in the forecast's confidence intervals. Yet, there are many issues needed to be further investigated. First, we should evaluate if the confidence intervals gained by the new method are accurate in real predictions and how they compare to the WRF model's intervals. This can be done by using a bigger data set and setting aside a proportion of the data set just for evaluating if the outputs really fall in the confidence intervals. Second, we should assess the significance of the method proposed when applied to different locations and seasons. Last but not least, more advanced statistical methods that can provide us with more accurate and dynamic (different width for different cases) confidence intervals can be applied to the problem.

References

[1]  M. Lange, On the Uncertainty of Wind Power Predictions-Analysis of the Forecast Accuracy and Statistical Distribution of Errors, J. Sol. Energy Eng. 127, 177 (2005)
[2]  T. H. Wonnacott, R. J. Wonnacott, Introductory Statistics, Wiley, New York, 1990
[3]  I. Jolliffe, D. B. Stephenson, Forecast Verification in Atmospheric Science: A Practitioner's Guide in Atmospheric Science, Wiley, 2002
[4]  M. Jorgensen, D. I. K. Sjoberg, An effort prediction interval approach based on the empirical distribution of previous estimation accuracy, Information and Software Technology, Volume 45, Issue 3, p123-136, 2003