TL;DR
This paper introduces an interval-valued neural network (IVNN) approach to quantify uncertainty in short-term wind speed predictions. By treating input data as intervals rather than single points, the IVNN can better capture inherent variability and provide more reliable prediction intervals (PIs) compared to traditional crisp methods.
What Problem Does This Paper Address?
Predicting future values of dynamic systems, like wind speed, is inherently challenging due to various sources of uncertainty. These uncertainties can stem from noisy or incomplete input data, imprecise model parameters, or limitations in the model's structure itself. Traditional prediction models often use single-valued (crisp) inputs and outputs, which can mask or misrepresent this uncertainty. This can lead to overly optimistic or unreliable predictions, especially in systems with significant inherent variability, such as wind speed, which fluctuates considerably within hours and days. The paper aims to address this by developing a method that explicitly models and quantifies this input uncertainty to produce more robust prediction intervals.
Methods
The core of the proposed method is an interval-valued neural network (IVNN). Unlike standard neural networks that process single numerical values, the IVNN is designed to accept and process interval-valued inputs. This means that instead of a single number, each input feature is represented by a lower and upper bound. The network then learns to map these input intervals to output intervals, which represent the prediction intervals (PIs) for the target variable (wind speed).
To train this IVNN and optimize the resulting PIs, the authors employ a multi-objective genetic algorithm, specifically the Non-dominated Sorting Genetic Algorithm-II (NSGA-II). The optimization process aims to simultaneously achieve two conflicting objectives:
- High Prediction Interval Coverage Probability (PICP): The proportion of actual values that fall within the predicted intervals.
- Low Prediction Interval Width (PIW): The narrowness of the predicted intervals, indicating a more precise prediction.
The NSGA-II algorithm explores a population of potential NN weight configurations, evaluating them based on these two objectives and iteratively refining the population to find a set of "Pareto optimal" solutions that represent the best trade-offs between PICP and PIW.
Datasets
The proposed method was evaluated on two types of datasets:
- Synthetic Case Study: Four time series datasets were generated using an Auto-Regressive Moving Average (ARMA) model with a 5-minute time frequency. These datasets were designed to simulate different input variability scenarios, using either Gaussian or Chi-squared distributions for the innovation terms, with varying variances.
- Real Case Study: Hourly wind speed measurements were used, collected over a three-year period (January 1, 2010, to December 30, 2012) from Regina, Saskatchewan, Canada.
For both case studies, the raw data was processed to create interval-valued hourly or daily inputs using two approaches: "min-max" (using the minimum and maximum values within an hour/day) and "mean" (using intervals defined by the mean and standard deviation). Crisp (single-valued) inputs were also generated as a baseline for comparison.
Key Insights and Ideas
- Interval Inputs Capture Real-World Variability: The fundamental insight is that treating input data as intervals, rather than single points, allows the model to inherently account for the uncertainty and variability present in real-world measurements (e.g., fluctuations within an hour or day).
- Multi-Objective Optimization is Crucial: Optimizing for both prediction interval coverage and width simultaneously is essential. A narrow interval with poor coverage is useless, and a wide interval with high coverage might not be precise enough. NSGA-II effectively navigates this trade-off.
- Crisp Methods Struggle with Variability: The comparison with crisp (single-valued) approaches clearly demonstrates that traditional methods are less reliable when dealing with significant input variability, often resulting in lower coverage probabilities.
- Interval-Valued Predictions Offer Better Decision Support: By providing a range of possible outcomes with a quantified probability, interval-valued predictions offer a more realistic and actionable basis for decision-making in critical applications like energy forecasting.
Key Results
- Improved Reliability: The interval-valued input approach consistently showed better performance in capturing input variability compared to crisp approaches, especially in scenarios with higher variability.
- Higher Coverage Probabilities: For the synthetic case study, the interval-valued approach achieved coverage probabilities that were generally higher and more stable than crisp methods, particularly when dealing with higher input variability.
- Comparable Interval Widths: While the primary focus was on coverage, the interval-valued approach managed to achieve comparable or even narrower prediction interval widths in some scenarios, indicating a good balance between precision and reliability.
- Validation of Generalization: The study demonstrated that the trained IVNN models generalized well to unseen data, with similar performance metrics observed on validation sets as on training sets.
Limitations and Future Work
- Subjective Solution Selection: The paper acknowledges that selecting a single "best" solution from the Pareto front is often subjective. While they propose a compromise, more formal methods for preference assignment could be explored.
- Computational Complexity: Multi-objective optimization with genetic algorithms can be computationally intensive, especially for large datasets or complex NN architectures.
- Future Work: The authors suggest exploring more advanced interval arithmetic techniques and investigating the application of this IVNN approach to other time series forecasting problems where uncertainty quantification is critical. They also highlight the potential for using interval-valued outputs to represent different types of uncertainty beyond just input variability.
Summary updated