An Interval-Valued Neural Network Approach for Uncertainty Quantification in Short-Term Wind Speed Prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Neural networks (NNs) often face challenges with uncertain input data.
• Quantifying prediction uncertainty from both data and models is crucial for reliable forecasting.
• Interval-valued data requires specialized methods for accurate prediction.

Purpose of the Study:

• To develop and demonstrate a neural network method for predictions with interval-valued input data.
• To quantify prediction uncertainty arising from both input data and the prediction model.
• To optimize prediction intervals (PIs) for accuracy (coverage probability) and dimension (width).

Main Methods:

• A multilayer perceptron NN was trained to map interval-valued inputs to interval outputs (PIs).
• Nondominated sorting genetic algorithm-II was employed for NN training, optimizing PIs.
• The method was validated using synthetic (autoregressive moving average model) and real-world (wind speed) datasets.

Main Results:

• The interval-valued input approach demonstrated superior reliability in capturing input variability compared to crisp methods.
• NNs trained with interval data produced prediction intervals (PIs) that better reflected real target value variability.
• Optimization using NSGA-II balanced PI accuracy and width effectively.

Conclusions:

• Neural networks can effectively handle interval-valued input data for prediction tasks.
• The proposed interval-valued input approach enhances prediction reliability by accounting for input uncertainty.
• This method offers a more robust alternative to crisp (single-valued) prediction approaches when dealing with uncertain data.