Intuitive lower upper bound estimation for short-term wind speed interval prediction

Estimating the uncertainty associated with wind speed forecasts is crucial for evaluating wind farms' power quality and operation. Prediction Intervals can communicate the uncertainty by prescribing a range of bounds within which the real value is expected to lie. While many approaches for PI generation exist, direct PI generation approach is usually preferred due to its simplicity. In this article we proposed a novel loss function to train neural networks to generate prediction intervals directly at its output. Building upon the fundamentals of the two state-of-the-art approaches; lower upper bound estimation (LUBE), and quality driven (QD), we proposed intuitive lower upper bound estimation loss (ILUBE). Utilizing the same objective of heuristic optimization of the two PI characteristics, coverage and interval size between, we applied a dynamic approach to obtain each of these two characteristics. The proposed ILUBE loss shows its strength in computational experiments involving eight different datasets for two different wind fields. The PIs generated by the proposed approach are tighter while covering a greater number of observations, in comparison to both LUBE and QD. Thus, by improving the PI quality, while keeping the essence of simplicity of the direct PI generation approaches, the proposed ILUBE loss emerged as a strong contender in this category.